Embed Size (px)
Citation preview
LAPPEENRANTA UNIVERSITY OF TECHNOLOGY
LUT School of Energy Systems
Energy Technology
Juha Juselius
ADVANCED CONDITION MONITORING METHODS IN
THERMAL POWER PLANTS
Master’s Thesis
Supervisor: D.Sc.(Tech) Jouni Ritvanen
ABSTRACT
Lappeenranta University of Technology
LUT School of Energy Systems
Energy Technology
Juha Juselius
Advanced Condition Monitoring Methods in Thermal Power Plants
Master’s thesis
2018
77 pages, 27 figures and 9 tables
Examiners: D.Sc.(Tech) Jouni Ritvanen
Professor, D.Sc.(Tech) Timo Hyppänen
Instructor: M.Sc. (Eng.) Ilkka Salmensaari
Keywords: condition monitoring, power plant maintenance, artificial neural networks
Power plant maintenance has been traditionally organized according to corrective and pre-
ventive maintenance strategies, which both have their limitations that can lead to high costs
due to low availability and inefficiently timed maintenance actions. Condition based mainte-
nance is considered an advanced maintenance strategy and is already applied in many critical
components in modern power plants. Extending condition monitoring and more efficient
maintenance practices to a wider range of equipment is desirable but requires advanced con-
dition monitoring methods that can detect changes in process parameters indicating faults
and malfunctions. These methods include different mathematical models that can be classi-
fied as physics-based, data-driven and hybrid models.
This thesis presents the theoretical background of different modelling approaches in condi-
tion monitoring applications. The theoretical part of this work also presents different mainte-
nance strategies and the role of condition monitoring in power plant maintenance. The main
objective of this thesis is to evaluate a commercial condition monitoring software based on
recursive neural networks, called Intelligent Health Monitor, in a case study where boiler’s
increased flue gas temperature is monitored. 96 models with different training parameters
are created and evaluated during training, healthy state monitoring and fault detection
phases. The results showed that reliable models with good fault detection capability can be
trained with automatically chosen input variables based on principal component analysis and
correlation analysis. The results also showed great variation and underlined the importance
of model verification. Based on the case study results and literature, several aspects that need
to be considered when applying artificial neural network based condition monitoring soft-
ware are discussed.
TIIVISTELMÄ
Lappeenrannan teknillinen yliopisto
LUT School of Energy Systems
Energiatekniikan koulutusohjelma
Juha Juselius
Kehittyneet kunnonvalvontamenetelmät lämpövoimalaitoksissa
Diplomityö
2018
77 sivua, 27 kuvaa ja 9 taulukkoa
Tarkastajat: TkT Jouni Ritvanen
Professori Timo Hyppänen
Ohjaaja: DI Ilkka Salmensaari
Hakusanat: kunnonvalvonta, voimalaitoksen kunnossapito, neuroverkot
Voimalaitosten kunnossapito on perinteisesti noudattanut korjaavia tai ennakoivia kunnos-
sapitostrategioita. Perinteisten strategioiden puutteiden ja niistä aiheutuvien korkeiden kus-
tannusten vuoksi kuntoon ja kunnonvalvontaan perustuvia kunnossapitostrategioita tarvi-
taan. Monien kriittiseksi luokiteltujen komponenttien kunnossapito perustuu jo kunnonval-
vonnan tuloksena saatuun komponentin kunnon arvioon, mutta kunnonvalvonnan katta-
vuutta tulisi laajentaa myös ei-kriittiseksi luokiteltuihin komponentteihin. Tämä edellyttää
kehittyneitä kunnonvalvontamenetelmiä, kuten matemaattisia mallinnusmenetelmiä, joilla
pystytään havaitsemaan prosessidatasta poikkeavuuksia, jotka indikoivat vikatiloja. Kun-
nonvalvonnassa käytetyt mallit voidaan luokitella fysikaalisiin malleihin, data-pohjaisiin
malleihin ja hybridimalleihin.
Tämän työn teoriaosassa esitellään eri kunnossapitostrategiat, kunnonvalvonnan merkitys
voimalaitosten kunnossapidossa sekä eri mallinnusmenetelmien teoreettinen tausta. Työn
päätavoitteena on testata ja arvioida kaupallista neuroverkkoihin perustuvaa kunnonvalvon-
taohjelmistoa Intelligent Health Monitoria. Ohjelmiston soveltuvuutta arvioidaan case-ta-
pauksessa, jossa voimalaitoskattilan savukaasun lämpötila on noussut laskien kattilahyöty-
suhdetta. Savukaasun lämpötilaa seurataan 96 mallilla, joiden tarkkuus arvioidaan kolmena
ajanjaksona: koulutusjakso, vikaantumattoman tilan jakso sekä vikatilanteen jakso. Tulokset
osoittavat, että luotettava neuroverkkopohjainen malli voidaan luoda käyttämällä automaat-
tisesti valittuja itsenäisiä muuttujia. Tulokset kuitenkin vaihtelivat huomattavasti mallien vä-
lillä, mikä korostaa käytettävien mallien verifioinnin tarpeellisuuden. Case-tapauksen tulos-
ten sekä kirjallisuuden perusteella nostetaan esiin useita näkökulmia, jotka tulee huomioida
käytettäessä neuroverkkopohjaista ohjelmistoa kunnonvalvonnassa.
FOREWORD
This thesis was done for Fortum Power Solutions as an independent study as part of Fortum's
contribution to ECSEL Joint Undertaking MANTIS. Firstly, I would like to thank Fortum
Power Solutions for providing me this great opportunity to work on this interesting topic and
exciting work environment. I want to express my gratitude towards all the Fortum’s experts
who have helped and guided me throughout this thesis. Especially I want to thank Ilkka
Salmensaari, Matti Kaija and Matti Visuri for all their valuable guidance and feedback. I
also want to thank my supervisor Jouni Ritvanen for all the provided instructions and com-
ments.
Last, but certainly not least, I wish to thank my parents and family for their unconditional
support from the beginning of my studies. Naturally, these special thanks are extended to
my beloved graphic designer.
Espoo, 31st of March 2018
Juha Juselius
TABLE OF CONTENTS
1 INTRODUCTION ............................................................................................... 6
1.1 Research goal .......................................................................................................... 9 1.2 Research structure ................................................................................................. 11 1.3 Fortum Power Solutions ........................................................................................ 11
2 POWER PLANT MAINTENANCE ................................................................ 13
2.1 Maintenance terminology ...................................................................................... 13 2.2 Maintenance strategies .......................................................................................... 15 2.2.1 Corrective Maintenance ..................................................................................... 15
2.2.2 Predetermined Preventive Maintenance ............................................................ 16
2.2.3 Condition based maintenance ............................................................................ 17
2.3 Characteristics in power plant maintenance .......................................................... 18
3 CONDITION MONITORING ......................................................................... 21
3.1 Fault detection, diagnosis and prognosis .............................................................. 22 3.2 Condition monitoring measurement techniques.................................................... 24
3.2.1 Vibration monitoring ......................................................................................... 24 3.2.2 Lubricant analysis and wear particle analysis ................................................... 24
3.2.3 Non-destructive deterioration monitoring ......................................................... 25 3.2.4 Performance monitoring .................................................................................... 25
4 DATA PREPROCESSING ............................................................................... 26
4.1 Data cleaning ......................................................................................................... 27
4.2 Sampling ............................................................................................................... 28 4.3 Data transformation ............................................................................................... 29
4.4 Feature selection/extraction .................................................................................. 29
5 MODELLING METHODS OF CONDITION MONITORING ................... 30
5.1 Physics-based modeling ........................................................................................ 32 5.2 Data-driven modelling .......................................................................................... 35 5.2.1 Statistical models ............................................................................................... 37
5.2.2 Artificial neural networks .................................................................................. 41 5.3 Hybrid modelling .................................................................................................. 47
6 CASE STUDY: DETECTION OF ABNORMALITIES WITH ARTIFICIAL
NEURAL NETWORK BASED SOFTWARE ................................................................ 48
6.1 Intelligent Health Monitor ..................................................................................... 49
6.2 Case study: Decreased boiler efficiency ............................................................... 49 6.2.1 Problem description ........................................................................................... 50 6.2.2 Evaluation criteria ............................................................................................. 54
6.2.3 Results ............................................................................................................... 57
7 ARTIFICIAL NEURAL NETWORK BASED CONDITION MONITORING
63
7.1 ANN training ......................................................................................................... 63
7.1.1 Data requirements and training period .............................................................. 64 7.1.2 Input variable selection ...................................................................................... 66
7.1.3 Network architecture ......................................................................................... 68 7.2 Model verification ................................................................................................. 69
7.3 Model training and verification summary ............................................................. 72
8 CONCLUSIONS ................................................................................................ 74
8.1 Future work ........................................................................................................... 77
REFERENCES
3
NOMENCLATURE
𝐄 residual matrix [-]
𝐅 residual matrix [-]
𝑓 measured value [-]
𝐈𝑚 identity matrix [-]
𝐽 performance index [-]
𝑘 discrete time [-]
𝐋 diagonal matrix [-]
𝐿 number of hidden layer [-]
𝑙 number of retained principal components [-]
𝑚 number of process variables [-]
𝑁ℎ number of hidden neurons [-]
𝑁 input-target relation [-]
𝑛 number of observations [-]
𝑛𝑖 number of inputs [-]
𝑛𝑜 number of outputs [-]
𝑛𝑝 number of input samples [-]
𝑜 output of a single neuron [-]
𝐏 matrix of orthogonal vectors [-]
�̂� matrix of retained eigenvectors [-]
�̃� matrix of ignored eigenvectors [-]
𝒑 input loading vector [-]
𝑄 heat energy [W]
𝒒 output loading vector [-]
𝐒 input score matrix [-]
𝑆𝑆 sum of squares [-]
𝒔 input score vector [-]
𝑇 temperature [°C]
𝑇f time to failure [s]
4
𝐓 matrix of transformed variables [-]
�̂� matrix of retained principal components [-]
�̃� matrix of ignored principal components [-]
𝑡 equipment current age [s]
𝐔 output score matrix [-]
𝒖 input vector [-]
𝐖 weight coefficient matrix [-]
𝒘 weight coefficient vector [-]
𝐗 input data matrix [-]
�̂� matrix of modeled variation [-]
𝐘 output matrix [-]
𝑌(𝑡) past condition profile up to current time [-]
𝒚 output vector [-]
𝑦 predicted value [-]
𝑧 target for single neuron [-]
Subscripts
abs absorbed
i input
reg regression
res residual
tot total
Greek Letters
𝛼 learning rate [-]
𝜂 efficiency [-]
𝜎 vector valued activation function [-]
Abbreviations
AI artificial intelligence
ANN artificial neural network
Auto automatic
5
CBM condition based maintenance
CI computational intelligence
CM condition monitoring
CMS condition monitoring system
ECSEL Electronic Components and Systems for European Leadership
FDI fault detection and isolation
FLS fuzzy logic system
GRNN generalized regression neural network
IHM Intelligent Health Monitor
IVS input variable selection
MAE mean absolute error
Man manual
MSE mean squared error
MIV mean impact value
MLP multilayer perceptron
O&M operation and maintenance
OSA-CBM Open Standard Architecture Condition Based Maintenance
PCA principal component analysis
PLS partial least squares
PoF physics-of-failure
PPM predetermined preventive maintenance
RBF radial basis function
RMSE root mean squared error
RUL remaining useful life
SOM self-organizing map
6
1 INTRODUCTION
The majority of the world's energy consumption is covered with conventional energy pro-
duction methods, such as hydro power turbines, nuclear fission, and combustion of gas, oil
or coal, see Figure 1. In power production, the actual production process is often optimized,
i.e. the efficiency is relatively high and not easy to increase without high investment costs.
At the same time, maintenance and availability are identified as crucial part of successful
operation. Low availability can lead to major costs in the form of lost production hours and
expensive repair work and therefore, the most important goal of a power operator is to max-
imize plant’s availability and performance with minimum costs (Kaija 2016, 1). However,
the production process and maintenance cannot be separated. Running the energy production
process inefficiently can cause significant costs if the production equipment would suffer
from wear and faults as the result of misuse. Consequently, proper information regarding the
production process and condition of the production equipment is identified as a vital part of
process development.
Figure 1. The world's comparative primary energy consumption in 2015 (The World Energy Council 2016,
4).
32.9 %
29.2 %
23.9 %
4.4 % 6.8 %
1.4 % 0.5 %0.9 %
Oil Coal Gas Nuclear Hydro Wind Solar Other renewables
7
Traditionally, industrial maintenance has been organized according to two strategies, namely
corrective and preventive maintenance. In corrective maintenance, actions are taken to re-
store equipment back to operational state after failures, whereas preventive maintenance
aims to prevent faults and malfunctions with planned maintenance actions before faults oc-
cur. Preventive maintenance actions, such as repairs and overhauls, are scheduled in advance
usually based on time or hours run. A fixed schedule for maintenance actions is often inef-
ficient, as maintenance actions can be carried out without actual need, or too late when
equipment performance has already degraded. To minimize the limitations and resulting
high costs of traditional maintenance strategies, see Figure 2, maintenance strategies based
on equipment's condition are required. As a result, the focus of maintenance strategies has
shifted from corrective and scheduled preventive maintenance to condition based mainte-
nance (CBM). This change has come from both technological developments and from
maintenance ideology transition. Industrial maintenance ideology has evolved from non-is-
sue "necessary evil" to strategic concern creating value to different stakeholders. Power sec-
tor makes no exception and recognizes CBM as the most attractive maintenance approach.
(Kobbacy & Murthy, 2008, 21; Martha de Souza 2012, 129)
Figure 2. Schematic illustration of maintenance cost distribution between maintenance strategies. Adopted
from Akula et al. (2017, 59).
8
Continuous operation, lost production during outages and very expensive capital equipment
leading to limited redundancy are characteristic to power plants. In general, commercial
earning of a power generator is dependent on the grid availability and operating efficiency,
and therefore, availability is often more valued criterion than reliability. (Chanda & Mukho-
paddhyay, 2016, 4). Power plants usually follow a combination of corrective and preventive
maintenance strategies augmented with predictive maintenance actions, meaning that the
equipment condition is predicted in a future state, on some critical components (Mobley
2002, 2–4). With the rapid development of sensor and data acquisition technologies, signal
processing and computational tools, it is desirable to extend condition monitoring to cover a
wider range of process equipment to enhance power plant’s availability and feasibility.
CBM is premised on condition monitoring activities to determine the physical state of a
component of a system. The condition assessment is used to schedule preventive mainte-
nance actions or to change operation mode to decrease unwanted degradation and wear of
equipment. The Open Standard Architecture Condition Based Maintenance (OSA-CBM) is
a widely used framework for industrial CBM. The framework, illustrated in Figure 3, is
designed by MIMOSA, which is an organization involved in CBM standard development
(Shin & Jun, 2015, 123).
Figure 3. The OSA-CBM framework (Thurston & Lebold, 2001).
9
The framework includes seven layers describing the whole CBM process from data acquisi-
tion to presenting the results. The first layer, i.e. sensor module consists of data acquisition,
which includes measurement devices and transducers for signal conversion. The second
layer, signal processing, is referred as data preprocessing in this thesis. This layer includes
several techniques, such as filtering and sampling, to enhance the data's reliability and to
transform the raw data into a more convenient format for further processing. Preprocessing
can be integrated to data acquisition with intelligent sensors or it can be treated separately
with additional modules. Condition monitoring layer gathers the preprocessed data to be
compared with predefined features. Diagnostic layer uses condition monitoring data to ob-
tain a health assessment of the monitored item. Prognostic layer predicts the item’s health in
a future state based on the results of condition monitoring and diagnostic layers. Execution
of maintenance actions is decided on decision support layer, which also considers available
recourses, such as spares, logistics and manning. The final layer, namely presentation, in-
cludes the man/machine interface used to display data and information to the user. This layer
may query all other layers. (Niu et al. 2010, 788; Shin & Jun, 2015, 123, Bousdekis et al.
2015, 1230)
Traditionally, condition monitoring is applied by setting fixed threshold values to monitored
parameters, and deviations from the set thresholds triggers an alarm. However, selecting
appropriate alarm limits is often challenging, and relies on human experience and
knowledge. Advanced condition monitoring methods, as understood in this thesis, are com-
putational models that detect abnormalities automatically without human-effort and indicate
the health of the monitored equipment with minimal human-aid. These methods are used in
the condition monitoring layer, diagnostic layer and prognostic layer.
1.1 Research goal
The research goal of this thesis is to study advanced condition monitoring methods that in-
clude computational models used as part of CBM. The aim is to study methods and tools
that are not limited to certain critical equipment, but instead allow plant-wide monitoring.
The focus is on computational models used to detect abnormalities, while other modules of
10
condition monitoring systems (CMS), such as sensor technology and data transfer, are out-
side the scope of this thesis. The following hierarchical terminology is adopted and used in
this thesis:
Approach is a combination of principles and methods defining a broad di-
rection in solving engineering problems. Examples of approaches in the con-
text of this thesis are physics-based modeling and data-driven modeling.
Methods or their combinations compose the chosen approach. A method is a
more targeted way than approach of describing how the problem is solved.
Examples of methods in the context of this thesis are statistical models and
artificial neural networks.
Techniques and tools form the solution method on a concrete level describ-
ing the precise strategy used. Examples of techniques and tools in the context
of this thesis are principal component analysis in the case of statistical meth-
ods, and multilayer perceptrons describing the neural network architecture.
The main objective of this thesis is to evaluate a neural network based commercial condition
monitoring software Intelligent Health Monitor by Algorithmica technologies Gmbh. (Al-
gorithmica technologies GmbH, 2018). The software is evaluated in a case study on a com-
bined cycle gas turbine power plant. This thesis presents one of the case studies, in which a
boiler’s gradually increased flue gas temperature is monitored. Based on the case study re-
sults and literature, several aspects that need to be considered when applying artificial neural
network based condition monitoring software are discussed answering the research question:
How to utilize artificial neural network based condition monitoring soft-
ware to support thermal power plant operation and maintenance?
Several advanced condition monitoring methods are applied to most critical power plant
components. Typically, CBM is applied to rotating machinery, such as steam and gas tur-
11
bines, pumps and compressors and electric motors and generators. Other common applica-
tions are bearings of different machinery, piping systems, heat exchangers and conveyor
systems (Kaija 2016, 13–19). Research related to condition monitoring methods has mostly
focused on fault diagnosis and prognosis, which both require clear differentiation between
different faults. Therefore, the research has also mainly focused on specific components with
well-known fault mechanisms. Numerous such applications have been well reviewed in the
literature. For instance, Henao et al. (2014) review diagnostics techniques of rotating elec-
trical machines, Leite et al. (2018) review prognostic methods of wind turbine maintenance,
Tahan et al. (2017) review diagnostics and prognostics methods of gas turbines and El-Thalji
& Jantunen (2015) review prognostics methods of rolling element bearings.
1.2 Research structure
The theoretical framework of this thesis consists of three sections. Section 2 presents the
common maintenance strategies used in thermal power plants. In Section 3, condition mon-
itoring as a crucial part of advanced maintenance is presented. Stages and objectives of con-
dition monitoring and commonly used condition monitoring measurement techniques are
introduced. Because of the enormous breath of maintenance as a research field, the objective
of these sections is not to provide in depth analysis of maintenance or condition monitoring,
but merely to introduce these topics. Section 4 gives an overview of some standard data
preprocessing techniques. In Section 5, three different modeling methods used in condition
monitoring are presented and discussed. Physical process models, data-driven models and
hybrid models are addressed on theoretical level. Section 6 presents a case study where in-
creased flue gas temperature is monitored using commercial condition monitoring software
Intelligent Health Monitor (Algorithmica technologies GmbH, 2018). In Section 7, practical
aspects of using artificial neural network based condition monitoring as part of a condition
monitoring system are discussed. Finally, conclusions are presented, and future work pro-
posed in Section 8.
1.3 Fortum Power Solutions
Fortum is a Finnish energy company with over 8000 (2016) personnel worldwide and a rev-
enue of 3600 M€ (2016). The Power Solutions unit offers operation and maintenance (O&M)
and related expert services and solutions to owners and developers of power and heat assets
12
while employing some 300 personnel (2016). The customer base of the Power Solutions unit
consists of both Fortum owned or co-owned plants and completely independent power plant
owners around the world. The main service areas are O&M services, burners and emission
control systems, power plant performance improvements, turbine and generator services and
IT-systems. (Fortum 2017a; Fortum 2017b)
The Power Solutions O&M unit operates and maintains power plants that are not owned by
Fortum. The O&M services utilize the O&M principles developed for Fortum owned and
serviced plants to offer O&M scheme optimization and audits as a service, including criti-
cality analysis, personnel training and strategy reviews. Most of the O&M operations pro-
vided by Fortum are based on the TOPgen concept, or its adaptation. The concept includes
setting up an efficient O&M organization, management systems and selected IT tools to
maintain the plant availability and performance at an optimal level during the whole life
cycle of the plant. (Fortum 2017b)
The performance service unit provides services to improve technical and economic perfor-
mance of power plants. The services include process analysis and modeling, plant operation
monitoring and optimization work delivery. The Turbines and Generators unit offers mainte-
nance and condition management services for power plant turbines in the form of mainte-
nance, inspections, condition management and consultation. The IT Systems unit provides
multiple tools to integrate in the power industry services. The four main products are TOPi
process monitoring, Solvo process simulation, Apros dynamic process simulation and Max-
imo asset management software. (Ibid)
This thesis is done for Fortum Power Solutions as an independent study as part of Fortum's
contribution to ECSEL Joint Undertaking MANTIS. 47 European industrial and academic
partners are contributing to the MANTIS program whose primary objective is to develop a
Cyber Physical System based Pro-active Maintenance Service Platform Architecture ena-
bling Collaborative Maintenance Ecosystems (MANTIS 2015).
13
2 POWER PLANT MAINTENANCE
European Standard EN 13306 (2001) defines maintenance as the combination of all tech-
nical, administrative and managerial actions during the life cycle of an item intended to retain
it in, or restore it to, a state in which it can perform the required function. The role and
importance of maintenance can be extended, and maintenance can be seen contributing not
only to profitable business, but to the aim of sustainable development in society, including
environmental, energy saving, safety, and economical aspects (Holmberg et al. 2010, 5).
Regardless of the industry, there is a consensus that maintenance is much more than a purely
technical matter and that a proper maintenance program is vital to attain business, environ-
mental and safety requirements. (Kobbacy & Murthy, 2008, 26; Al-Turki et al. 2014, 5)
Maintenance consists of a vast set of problems and operational implications, which makes
the various approaches to maintenance diverse and difficult to schematize in a general man-
ner. Traditionally, maintenance is classified into corrective and preventive maintenance. Be-
sides types of maintenance actions, also multiple strategic and philosophical approaches to
maintenance, such as e.g. total productive maintenance that focuses on constant development
of the overall equipment effectiveness (Wireman 2004, 14) and reliability centered mainte-
nance that focuses on determining the most effective maintenance approach for individual
equipment (Moubray 1997, 7–8), have been developed and implemented in different indus-
tries. (Holmberg et al. 2010, 1; Fedele 2011, 33)
2.1 Maintenance terminology
Fault and failure are two fundamental terms describing the state of a system or a component.
Fault is considered as an abnormal change of system behavior, which deteriorates the per-
formance of the system. Faults can be classified as illustrated in Figure 4. Abrupt faults occur
rapidly, when the parameter indicating faulty behavior abruptly rises to a new constant level.
Incipient faults develop gradually over time, for example, in the case of slow degradation of
a component. Intermittent faults have a repeated pattern, i.e. they occur and disappear re-
peatedly. A failure is a consequence of progressing incipient fault over time demolishing the
normal operation of the system (Sobhani-Tehrani & Khorasani, 2009, 1–2). (Patan 2008, 9)
14
Figure 4. Fault classification: abrupt (dashed), incipient (solid) and intermittent (dash-dot) (Patan 2008, 9).
Maintenance is still a relatively young branch of science, and while some terms have unam-
biguous definitions, there is still a lot of confusion with maintenance terminology both in
literature and in practice. Many terms are used differently between sources and misappre-
hensions are not uncommon. (Kobbacy & Murthy, 2008) To avoid any confusion, the fol-
lowing elementary maintenance terms are presented and adopted in this thesis.
Maintenance policy indicates the overall attitude assumed in relation to maintenance prob-
lems. Examples of maintenance policies are total productive maintenance and reliability cen-
tered maintenance. Maintenance policy can be clarified in the use of various strategies.
Maintenance strategy follows maintenance policy in hierarchy. Maintenance strategy is an
operational approach to maintenance problems based on the criteria provided by the adopted
maintenance policy. Maintenance strategy determines the triggering mechanism for mainte-
nance actions. Examples of maintenance strategies are preventive and predictive mainte-
nance strategies.
Maintenance actions are basic maintenance interventions and tasks performed by a techni-
cian. Examples of maintenance actions are inspections, repairs and overhauls.
15
As stated above, literature still lacks unambiguous terminology and a lot of confusion exists.
For instance, Kobbacy & Murthy (2008), Fedele (2011) and Al-Turki et al. (2014) all give
different definitions to maintenance terms.
2.2 Maintenance strategies
According to EN 13306 (2001), maintenance can be classified into corrective maintenance
and preventive maintenance. Both maintenance types can be classified further into sub-cat-
egories, as presented in Figure 5.
Figure 5. Maintenance types according to standard EN 13306:2001 (Crespo 2007, 70).
2.2.1 Corrective Maintenance
Corrective maintenance, also known as run-to-failure or breakdown maintenance, is carried
out to restore or repair an equipment or a system after a fault back to a stage in which it can
perform a required function. Corrective maintenance can be immediate or referred based on
the timing of maintenance actions. Immediate maintenance is performed without delay after
fault detection to avoid unacceptable consequences, whereas deferred maintenance is de-
layed after fault detection according to given maintenance rules. Figure 6 presents the typical
activities of corrective maintenance tasks. (Crespo 2007, 71)
16
Figure 6. Flowchart of activities of a corrective maintenance task (Martha de Souza 2012, 128).
As corrective maintenance is reactive by nature, it is only appropriate approach if the failure
consequences are small. The more complex and critical the component is for the process, the
more severe consequences a fault may cause. If applied correctly, corrective maintenance
has some advantages, as maintenance planning is simple, and no work is scheduled until
needed. Also, some failures are not predictable by any instruments or analysis and manual
inspection as part of corrective maintenance is the only way to detect a fault. (Holmberg et
al. 2010, 11)
Despite the simplicity and advantages of corrective maintenance, there are some major dis-
advantages when applied in inappropriate situations. For instance, failures can occur at an
inconvenient time leading to loss of production, some faults may go unnoticed causing ex-
pensive consequential damage, a large maintenance crew and spares inventory is needed on
standby and there is no data available regarding the past, present or possible future state of
the equipment. In general, expensive and even dangerous failure consequences should be
expected when corrective maintenance is applied inappropriately. (Ibid)
2.2.2 Predetermined Preventive Maintenance
Preventive maintenance aims to maintain equipment in working condition by performing
maintenance actions before failures occur to minimize the failure consequences related to
corrective maintenance. Preventive maintenance can be predetermined, or condition based.
(Crespo 2007, 70–71)
17
In predetermined preventive maintenance (PPM), maintenance is scheduled in advance
based on e.g. time, hours run, operation cycles or production. PPM is used the most effec-
tively, when machine/component life is predictable, or age related. The most important ad-
vantage of PPM compared to corrective maintenance, is a more effective use of time, as
plant downtime and failure consequences are decreased and maintenance is planned well in
advance. PPM also increases machinery's reliability and reduces the need for spares. The
typical activities of PPM tasks are presented in Figure 7. (Chanda & Mukhopaddhyay, 2016,
98; Holmberg et al. 2010, 12)
Figure 7. Flowchart of activities of a PPM task (Martha de Souza 2012, 129).
The main disadvantage of PPM is that scheduled actions are often not carried out at optimal
time. Equipment may not need maintaining but due to scheduled maintenance actions, spares
and labor are used unnecessarily and potential plant downtime is caused. On the other hand,
faults may still occur, when components time to failure is shorter than the maintenance in-
terval. Optimizing the maintenance interval to minimize overall costs is crucial to successful
PPM. (Holmberg et al. 2010, 12)
2.2.3 Condition based maintenance
In condition based maintenance (CBM), component's health is monitored, and predictive
maintenance actions are carried out to prevent failures before they occur. CBM is considered
an advanced maintenance strategy combining the benefits of other strategies. The main ad-
vantage of CBM is that equipment can be maintained operational with minimum mainte-
nance. This means that better planning of repairs is possible and reduced spare inventory is
needed, inconvenient breakdowns and expensive consequential damage are avoided, and the
failure rate and unnecessary work are reduced. Studies have shown that implementing a
18
CBM strategy can reduce the number of component breakdowns by 70% to 75%, mainte-
nance costs by 25% to 30% and equipment downtime by 35% to 45% (Sullivan et al. 2010,
5.4). The typical activities of predetermined preventive tasks are presented in Figure 8.
(Holmberg et al. 2010, 15)
Figure 8. Flowchart of activities of a predictive maintenance task (Martha de Souza 2012, 130).
CBM requires condition monitoring to be effective. Measuring techniques used in condition
monitoring can vary from human sensing to sophisticated instrumentation aiming to detect
the current state or even to predict the future condition of the monitored component. Differ-
ent measurement techniques are presented in Section 3. The main challenges of CBM are
related to the task of condition monitoring. The required instrumentation and data processing
system can be expensive and accurate maintenance can be difficult to achieve due to com-
plexity of the system and its failure mechanisms. (Holmberg et al. 2010, 5; Chanda &
Mukhopaddhyay, 2016, 4; Niu et al. 2010, 786–787)
2.3 Characteristics in power plant maintenance
Section 2.2 has introduced the common industrial maintenance strategies summarized in Ta-
ble 1. Even though condition based maintenance has some superior features compared to
less advanced strategies, it is essential to emphasize that the best maintenance policy incor-
porates all mentioned strategies. Every strategy has their own strengths and limitations and
each strategy used appropriately is part of a planned approach. According to Hashemian &
Bean (2011), almost 30% of industrial equipment does not benefit from CBM. (Holmberg
et al. 2010, 16)
19
Table 1. Range of maintenance approaches. Adopted from Niu et al. (2010, 787).
Category Maintenance approaches
Sub-category Corrective Preventive
Run-to-failure Predetermined Predictive
Fix when it
breaks
Scheduled
maintenance
Condition
based
maintenance
diagnostics
Condition based
prognostics
When scheduled No scheduled
maintenance
Based on a fixed
time schedule
Based on
current
condition
Based on
forecasting of
remaining
equipment life
Why scheduled N/A Intolerable
failure effect and
possibility of
preventing the
failure effect
Based on
evidence of
needs
Maintenance
need is projected
as probable
within mission
time
How scheduled N/A Based on the
useful life of the
component
forecasted during
design and
updated trough
experience
Continuous
collection of
condition
monitoring
data
Forecasting of
remaining
equipment life
based on actual
stress loading
Kind of prediction None None On and off-
line, near real-
time trend
analysis
On and off-line,
real time trend
analysis
There are different models and guidelines to maintenance management and how to assign
maintenance strategies on different industrial fields, one popular approach being reliability
centered maintenance. The generally applied maintenance strategies are also used in thermal
power plants but compared to traditional maintenance practices, thermal power plants have
some characteristic features. First, power plants are most often run continuously, and mainte-
nance actions may cause outages of the plant operation. Duration of outages are restricted
by market demand and therefore, an appropriate maintenance philosophy with a short-term,
medium-term and long-term approach to the market demand pattern is required. Second,
outages during maintenance cause loss of production, which is often a major expense with
20
unplanned outages. An example of unplanned unavailability costs is presented in Figure 9.
Elaborated further, commercial earning of a power generator is dependent on the grid avail-
ability and operating efficiency, and therefore, availability is often more valued criterion
than reliability. Third characteristic feature of power plant maintenance is expensive main
capital equipment, which means that some key process equipment cannot be duplicated that
in turn leads to limited redundancy. (Chanda & Mukhopaddhyay, 2016, 4; Ndjenja & Visser,
2015, 136)
Figure 9. Unplanned unavailability costs in power plants. Adopted from Kylliäinen et al. (2011, 4).
Most modern power plants use a combination of corrective and preventive maintenance
strategies. The most effective strategy for individual items and components are determined
by several factors, e.g. the component's failure rate and criticality, which means that the
optimal combination of maintenance strategies is unique for a specific plant or system to
enhance the plant's availability at the lowest maintenance cost. (Chanda & Mukhopaddhyay,
2016, 4; Ndjenja & Visser, 2015, 136)
The use and continuing development of various diagnostic techniques, high system and com-
ponent automation level and more efficient work organizations have contributed to the better
optimization of maintenance practices in recent years. Despite continuous improvement and
development, a lot of potential for further cost reduction still exists, especially by extending
21
condition monitoring to equipment not rated critical. Minimizing maintenance costs is criti-
cal for power plant operators, since maintenance can cover up to 30% of all electricity gen-
erating costs. (Gräber 2004, 1–2)
3 CONDITION MONITORING
Condition monitoring is a process of monitoring parameters that provide information about
changes in the component's or system's condition or performance. Condition monitoring can
be applied online to assist the continuous operation or as an offline service to support longer
term operation. Condition monitoring includes all the actions from data acquisition to esti-
mating the state of health of the monitored component. A generalized condition monitoring
framework is illustrated in Figure 10, with each box presenting one device. The first box
contains data acquisition device, whose primary function is to acquire data from the system.
Data acquisition devices usually include measurement devices, such as thermometers, accel-
erometers or strain gauges. The second component is data preprocessing, which can be inte-
grated to data acquisition or treated separately. Often some preprocessing is integrated to
data acquisition to reduce the data transfer and storing requirements, and additional prepro-
cessing is performed with separate modules to enable efficient data analysis. Some data pre-
processing techniques are described in Section 4. The preprocessed data are processed fur-
ther in feature selection device, which is a process of identifying and quantifying specific
aspects of the data that are proper indicators of faults in the structure or process. In the deci-
sion making device, the selected features are interpreted to obtain condition diagnosis as the
outcome. (Marwala 2012, 4–5)
22
Figure 10. Condition monitoring framework. Adopted from Marwala (2012, 4).
3.1 Fault detection, diagnosis and prognosis
Three important stages in condition monitoring can be identified: fault detection, fault diag-
nosis and prognosis. Fault detection is a process to recognize abnormal operating conditions
indicating faults in the monitored system. Diagnosis in turn, is a process to analyze the ex-
istence and the cause of the problem, whose objective is to examine symptoms and syn-
dromes to determine the nature of faults and failures (ISO 13372, 2012). Diagnosis is com-
prised of fault isolation, which is a task to locate the faulty component, and of fault identifi-
cation, which is a task to determine the nature and severity of the fault. The definition of
diagnosis varies in the literature. Fault detection can be treated separately, as in this thesis,
or it can be seen as part of diagnosis. In practice, fault identification phase rarely appears,
and thus, the common abbreviation used in many scientific papers is FDI (Fault Detection
and Isolation) (Patan 2008 ,10). (Van Tung & Yang, 2009, 62; Jardine et al. 2006, 1485)
The final and most appealing stage of condition monitoring process is prognosis, which is a
process of predicting a future condition from present signs and symptoms. Prognosis aims
to predict faults before they occur determining the failure mode, when the fault will occur
and its chance of happening (Leturiondo 2016, 15). Prognosis predicts the component's time
23
to failure in terms of Remaining Useful Life (RUL) within design parameters. RUL, some-
times referred as residual service life, can be defined as the time left until the monitored unit
cannot perform its specified function given the current machine age and condition, and the
past operation profile. RUL can be expressed as conditional random variable as:
RUL = 𝑇f − 𝑡|𝑇f>𝑡,𝑌(𝑡), (1)
where 𝑇f denotes the random variable for time to failure, 𝑡 is the current age and 𝑌(𝑡) is the
past condition profile up to the current time. Being a random variable, the information ob-
tained from Eq. 1 can be divided into two characteristics: Estimation of the RUL and prob-
ability distribution of the RUL. When in healthy state, statistical estimations can be used to
calculate the RUL but when evidence of a failure occur, accurate information from the diag-
nosis process is required for the RUL estimation. (Leturiondo 2016, 14–18; Yan 2014, 153–
154; Pusey 2007, 3–4)
Stages of condition monitoring to support maintenance decision making are presented in
Figure 11. However, it should be highlighted that fault diagnosis and prognosis require clear
differentiation between different faults, which in turn requires a fault library of the possible
faults of the monitored item. Creating and sustaining a fault library is possible when moni-
toring a limited amount of equipment with known fault mechanisms. However, extending
the fault library to cover plant-wide monitoring is still practically impossible. In plant-wide
monitoring, the main aim of condition monitoring is still to detect fault development as early
as possible and assist human decision making.
Figure 11. Stages of condition monitoring to support maintenance decision making.
24
3.2 Condition monitoring measurement techniques
Several measurement techniques have been applied in condition monitoring aiming at
providing information of the actual condition of the monitored component or system. The
simplest form of condition monitoring are routine inspections carried out by plant personnel
to detect local abnormalities, such as noise, smell or vibration in the running component.
While the importance of inspections by experienced personnel cannot be neglected, more
accurate automatized and intelligent condition monitoring systems are desired, especially
for decentralized production and remote support services. Some of the most common sensing
techniques are vibration monitoring, lubricant analysis, wear particle analysis, monitoring
of performance parameters and deterioration monitoring using non-destructive techniques.
(Manning-Ohren 2015; Martha de Souza 2012, 130–131; Marwala 2012, 1)
3.2.1 Vibration monitoring
All running machines vibrate while running. Changes in vibration levels and patterns can be
used for detecting faults and identifying the health of an equipment. Vibration monitoring is
the most used measurement technique in condition monitoring, especially with rotating ma-
chines such as pumps, compressors, turbines and electrical generators and motors. Typical
measured parameters are acceleration, velocity and displacement of moving mechanical
components. These parameters can be measured directly at the component of interest but
more often vibration is measured at an available surface giving an indication of internal
events without disrupting the process. Vibration monitoring equipment consists of three
main items: transducers for data acquisition, signal processing system and an algorithm for
condition assessment. Typically, the raw vibration data is processed into overall vibration
levels, frequency spectra or high frequency emission. (Opetushallitus; Martha de Souza
2012, 131; Holmberg et al. 2010, 20)
3.2.2 Lubricant analysis and wear particle analysis
Lubricant analysis measures the additives, contaminants and debris, which reveal the condi-
tion of the lubricant. The analysis aims at scheduling lubricant change intervals optimally to
maintain satisfactory equipment operation. Lubricant analysis is often coupled with wear
particle analysis when samples of lubricant oil are collected and analyzed in pre-determined
intervals. Rate change of debris collection indicates changes in the system condition. Wear
25
particle analysis is also important to avoid high level of wear particles from rubbing surfaces
in the machinery. (Martha de Souza 2012, 131; Holmberg et al. 2010, 21)
3.2.3 Non-destructive deterioration monitoring
Several non-destructive techniques, which evaluate the properties of a material without af-
fecting the integrity of the component under test, are being used to monitor deterioration of
some components. Typically, non-destructive testing is used for static equipment, such as
pressure vessels, piping systems and structural components which undergo high stresses due
to high temperature and pressure or are prone to corrosion and erosion. Non-destructive
techniques allow inspections of difficult-to-reach hot spot stress areas without interfering the
process. It is also possible to define the importance of detected defects for structural integrity
to define the need for maintenance actions. The most used non-destructive techniques in-
clude magnetic particle inspection, eddy current inspection, acoustic emission testing, radi-
ographic inspection and ultrasonic inspection. Thermographic inspection, which measures
infrared energy emissions, is also considered a non-destructive technique but is usually used
to evaluate possible cumulative damage evolution in electrical equipment. (Martha de Souza
2012, 131; Urata et al. 2015, 37–39)
3.2.4 Performance monitoring
Monitoring the performance of an equipment is an important aspect of condition monitoring.
The performance, which can be inspected by monitoring performance or process parameters,
such as pressure, flow rate or energy consumption of an equipment, can indicate the condi-
tion of a machine. These indications can be used to diagnose the system operational condi-
tion. Failure condition can be recognized using fixed limit values or by detecting abnormal-
ities. Traditionally, experienced personnel are required to detect and identify failure modes
from process parameters, but the recent development of computational methods have led to
a point, where the need for human diagnosis and localized inspection techniques can be
greatly reduced using computational methods. Computational intelligence methods, such as
artificial neural networks, have gained a great deal of attention and led to the development
of global methods, which can use changes in measured data as a basis for fault detection
reducing human effort and detecting faults at early states invisible to traditional approaches.
(Marwala 2012, 1; Holmberg et al. 2010, 20; Martha de Souza 2012, 131)
26
4 DATA PREPROCESSING
Data acquisition, which is a process of collecting and storing useful data from the monitored
asset, is a crucial part of any condition monitoring system both for monitoring and modelling
purposes. The collected data under a CBM program can be categorized in two main types:
event data and condition monitoring data. Event data include the information of performed
maintenance actions, e.g. installations and repairs, and is usually added manually to the sys-
tem. Condition monitoring data consist of the data formed from measurements related to the
process or the condition of the component. The acquisition of condition monitoring data is
usually integrated to plant automation. The process data can be augmented with additional
instrumentation, e.g. portable devices, on the most critical components, but extending addi-
tional instrumentation to cover the whole plant is often too expensive for CBM purposes.
Therefore, methods capable of detecting fault development from process data without the
need for additional instrumentation are desirable to extend accurate condition monitoring to
cover as many equipment as possible. As described in Section 3.2, different measurements
are used in condition monitoring and therefore, condition monitoring data are very versatile.
The data can be, for example, vibration data, acoustic data, oil analysis data, temperature,
pressure or environment data. The instrumentation requirements also vary between machin-
ery. Some slowly changing parameters, e.g. temperature, can be measured with an interval
of seconds or even ten seconds, whereas fast parameters, e.g. vibration, are usually measured
with a period of milliseconds. The measurement of fast changing parameters can be then
averaged over a longer time period depending on the data analysis requirements. (Niu 2017,
50; Jardine et al. 2006, 1485; Kaija 2015, 25–26)
Data preprocessing includes any type of raw data processing to enhance the reliability of the
data, which in turn improves the accuracy of data analysis (Niu 2017, 71). Data prepro-
cessing methods can be broadly divided into two categories: data preparation and data re-
duction. Data preparation includes methods and techniques that are used to initialize raw
data in a way that the data can be effectively processed for the purpose of the user, for ex-
ample, in a modeling software or for expert analysis. Data reduction refers to a set of tech-
niques, which obtain a reduced representation of the data making analysis algorithms faster
and more feasible. Data reduction is not always mandatory for CMB purposes, but it can be
27
highly useful for machine learning algorithms and when using large amounts of data. (Garcia
et al. 2015, 11–13; Bangert 2012, 37–38)
The type and extent of required data preprocessing greatly varies between applications and
the chosen approach. Many methods also overlap and can serve several purposes, e.g. data
cleaning and transformation, at once. The most used methods include (Bengtsson et al. 2004;
Niu 2017, 71):
Data cleaning to remove noise from data,
Sampling to select a representative subset from a large population of data,
Transformation to manipulate raw data into a more convenient format
Feature selection/extraction to extract specified data significant in some particular
context.
4.1 Data cleaning
The first step of data preprocessing is data cleaning. Sensors need to operate under harsh
conditions in power plant environment, which may lead to unwanted effects and so-called
dirty data not representing the actual measured quantity. The cause for unreliable dirty data
can be, for example, broken or faulty sensor due to clogging, dirt or overheating. Also, in-
correct sensor installations, various temporal and programming effects and sources of noise
along the measurement chain can cause artificial variation to the measurements. Since event
data is often entered manually to the system, also event data contain errors due to human
factor. (Stefaniak et al. 2014, 200–203; Bangert 2012, 37–38; Jardine et al. 2006, 1485)
Outlier detection is one form of data smoothing and noise reduction carried out to capture
the studied dynamics. Outliers can be detected based on their type. Unrealistic data e.g. tem-
perature below 0 Kelvin are eliminated by introducing a minimum and maximum allowed
value for every sensor. Unlikely data, which are not necessarily unphysical but most likely
caused by faulty sensors can be detected with statistical tools, usually setting acceptance
limits in terms of number of standard deviations away from the mean. In contrast to fixed
limits, statistic parameters change over time creating dynamic limits to accepted values.
Third method to detect outliers is clustering. In this method, clusters of data points presenting
a particular operating condition are formed. Data points not belonging to any cluster are
28
irregular by definition and can be excluded from processing. Outlier detection is also a prob-
lem of missing data. Often missing data points can be replaced with artificial values, obtained
with interpolation or with soft sensors, to enable data analysis. Other possibility is to exclude
the missing data from the analysis, but this may be problematic to the analysis, especially if
the missing data have a structured cause. (Bangert 2012, 42)
As with other data preprocessing methods, also data cleaning methods vary between appli-
cations. One particularly important field is signal processing, which refers to data processing
for waveform data. Waveform data, e.g. vibration signals, are widely used in condition mon-
itoring applications. Several techniques to de-noise, i.e. to increase the signal-to-noise ratio,
waveform data are used, e.g. wavelet transform and different filters that remove portions of
the signal identified as noise. (Niu 2017, 70–71; Roulias et al. 2013, 1045–1046; Bangert
2012, 47)
4.2 Sampling
Sampling is a data reduction technique used to ease the data analysis of large data sets. For
data-driven modeling, sampling refers to the decision of which data points are presented to
the learning algorithm. Sampling also includes the decision, whether the data points are pre-
sented once (sampling without replacement) or multiple times (sampling with replacement)
to represent a certain condition. Besides reducing the amount of used data, data sampling is
also important to balance the data regarding the occurrence of rare events. Data sampling
can have several forms, random sampling and stratified sampling being the two basic tech-
niques. In random sampling, data points are chosen based on a probability distribution,
which is usually the uniform distribution resulting in equal probability for all data points.
Stratified sampling is used with imbalanced data sets selecting the sample data points based
on a probability distribution that results in over-sampling rare data points representing ab-
normal conditions, and in under-sampling common data points representing normal fre-
quently occurring conditions. (Bangert 2012, 50–51; Garcia et al. 2015, 156–158)
29
4.3 Data transformation
Data transformation is a process of converting raw data in a way that it can be applied more
efficiently (Garcia et al. 2015, 12). The type and extent of data transformation is highly de-
pendent on the data type and on the analysis method. For instance, vibration signals by de-
fault require more preprocessing, including wavelet transformation and Fourier transfor-
mation, in comparison to temperature data. Data transformation includes several methods,
e.g. smoothing, feature construction and aggregation, discretization and normalization being
the most important and used ones.
Discretization transforms continuous attributes into discrete ones (Garcia et al. 2015, 244).
In industrial setting, most of the sensor readings are identical. As storing every single meas-
ured value is often inefficient, the time-series from instrumentation is discretized and only
certain values accepted as sufficiently new are recorded. Usually the discretization is done
according to one of the three following criteria: a new value is recorded at a manually set
fixed time interval; a new value is recorded if it differs enough from the previously recorded
value or a new value is recorded at a fixed time interval, unless the value is outside certain
boundaries when the rate for value recording is adjusted to keep a detailed record of unusual
events. (Bangert 2012, 38–40)
Normalization refers to data transformation techniques that do not generate new attributes
but transform the distribution of the original values resulting in a new set of values. Normal-
ization is a standard data preparation method for analysis based on statistical methods and
machine learning. The most used normalization technique is the min-max normalization
scaling all numerical values to a specified range, usually from 0 to 1. Other common tech-
niques are Z-score normalization that uses standardization based on standard deviation, and
decimal scaling normalization that uses power of then division to transform a numerical
attribute resulting in absolute values lower than 1. (Garcia et al. 2015, 46–48).
4.4 Feature selection/extraction
Feature selection is a common technique for data and dimensionality reduction. Feature se-
lection is used to identify irrelevant or redundant features to enable more efficient classifi-
30
cation. Feature selection algorithms fall into three categories: filters are used to extract fea-
tures from the data without any learning involved; wrappers use learning techniques to find
useful features; and embedded techniques which combine feature selection and classifier
construction. (Hira & Gillies, 2015, 1–3; 2; Bania 2014, 3–4)
Feature extraction is also used in data reduction. The difference compared to feature selec-
tion is that feature extraction calculates new features from the original ones. In general, fea-
ture extraction is a mapping of a multidimensional space into a lower-dimensional subspace.
The most well-known technique for dimensionality reduction is principal component analy-
sis described in Section 5.2.1. (Hira & Gillies, 2015, 6–7)
5 MODELLING METHODS OF CONDITION MONITORING
Applying CBM strategy to thermal power plants requires system modelling and process
monitoring. The main approaches to system modelling are physics-based modelling based
on a prior knowledge, data-driven modelling, and hybrid modelling, which combines several
modelling techniques (Leturiondo 2016, 17). A classification of modeling methods in con-
dition monitoring is presented in Figure 12 and these approaches are discussed in this sec-
tion. Numerous models for fault detection, diagnosis and prognosis have been reported in
the literature to support maintenance decision making. As the field of condition monitoring
modeling is enormous, this section is not intended as a comprehensive list of modelling
methods but rather to provide an overview of different approaches. This section concentrates
on fault detection, since diagnosis and prognosis are still very difficult to apply in plant-wide
monitoring.
31
Figure 12. Classification of system modeling approaches in process and condition monitoring. Adopted from
Das et al. (2012, 722).
In the context of CBM, the main task of modelling is to detect abnormalities indicating faults
and malfunctions. In the simplest form, fixed limits are assigned to each monitored meas-
urement and alarms are created when measurements exceed the threshold. Although simple,
this approach has multiple limitations. Manually applied limits require a great deal of process
knowledge and engineering work but the limits are still mainly based on assumptions and
experience. Fixed limits are also difficult to set in a way that they are representative for
changing operation and environmental conditions. The main technique to overcome these
limitations is to use residuals as fault detector. Deviations between the monitored and ex-
pected values, called residuals, are artificial time-varying signals reflecting potential faults,
see Figure 13. In healthy state, the residual is normally designed to be zero and significantly
differ from zero when the monitored item is in faulty state. As real systems are always sub-
ject to noise and models possess some level of uncertainty, the residual value is never ex-
pected to be zero but rather in the neighborhood of zero. (Frisk 2001, 9–11; Venkatasubra-
manian et al. 2003a, 300–302)
32
Figure 13. Schematic illustration of fault detection using residuals. Adopted from Sheriff et al. (2017, 238).
The residual approach consists of residual generation and residual evaluation. Residual gen-
eration refers to the comparison of measured and predicted outputs obtained by the chosen
model, and residual evaluation to the analysis whether a fault has occurred including also
fault isolation (Patan 2008,10). Residual generation is considered a more important stage,
since residual evaluation is fairly simple if residuals are well-generated (Sobhani-Tehrani &
Khorasani, 2009, 12). In many applications, residual generation is enough for fault detection,
when residuals correspond directly to the equipment under abnormal behavior or the main
task of condition monitoring is merely to get the operator's attention (Stanley 2010). For
fault isolation and diagnosis, residual evaluation including a prior knowledge of possible
faults is required (Frisk 2001, 12).
5.1 Physics-based modeling
Physics-based modeling, also known as modeling by first principles, is based upon natural
laws providing a quantitative description of the system and its physical phenomena (Stringer
et al. 2012, 156). Physics-based models can be classified in various ways. For instance,
whether the model considers time variations, models are classified as static or dynamic or as
linear or non-linear based on the obtained set of equations. (Jaako 1998, 9–12)
Static models present the modeled process at a single point in time and cannot describe time
variations. These models are described by algebraic or partial different equations where the
time derivative term is neglected. This makes them simpler and faster to evaluate than dy-
33
namic models, but since real processes are hardly ever static, the applicability of static mod-
els is limited mostly to process design and to slowly changing processes, which can be ap-
proximated as non-time-dependent. Dynamic models, in contrast to static models, describe
time variations. These models are described by ordinary differential equations, differential
algebraic equations or partial differential equations making them more complex to solve than
static models. (Jaako 1998, 10–11; Ge 2017, 21)
One classification of physics-based models used in condition monitoring applications is to
classify the models as system models or physics-of-failure (PoF) models. System models
rely on residuals, whereas the PoF approach uses damage models to determine fault genera-
tion and propagation (Pecht & Gu, 2009, 310). This approach requires an accurate analytical
model, which describes how variables like material properties, geometry, load cycles and
operating conditions impact on different failure modes and degradation mechanisms. PoF
models are widely used in electronic systems (Temsamani et al. 2017). In power generation,
they are mostly implemented to components, such as piping systems (Di Maio et al. 2015;
Chookah et al. 2011) and turbine discs (Zhu et al. 2016), object to creep, fatigue, erosion and
corrosion. These models are highly specific to each component and application and are
mainly used in reliability assessment, but they can also be implemented into more fast-paced
condition monitoring to provide continuously updated predictions of the structural health of
some critical components (Tautz-Weinert & Watson, 2017). (Pecht & Jaai, 2010, 318; Mar-
tha de Souza 2012, 261–262)
Another classification of physics-based models used in condition monitoring is based on
residual generation and evaluation approaches, as presented in Figure 12. The most used
techniques fall into three categories: parameter estimation based techniques, observer based
techniques and parity equation based techniques. The main idea of comparing measured and
predicted values is the same for all three approaches, but the mathematical presentations of
predicted values differ.
Parameter estimation techniques rely on detecting changes in the physical process parame-
ters not considered in the development of the model. A physical process parameter can be,
for example, heat transfer coefficient in the case of heat exchangers. The model parameters
34
are functionally related to the physical parameters and this relationship in the form of math-
ematical equations is used to estimate the physical parameters. The estimated physical pa-
rameters can then be compared to measured values. Observer, or filter, based approach mon-
itors a process with state variables. Observer based methods compare measurements to
model based reconstructions of the system outputs that are estimated by different observers
and filters, e.g. Kalman filters. Parity equations are rearranged and often transformed vari-
ants of the input–output or state–space model of the considered process. Parity equations are
used to estimate the residual values. Rearranging the model structure enables efficient fault
isolation by generating residual vectors orthogonal to each other for different faults. (Das et
al. 2012, 734–736; Tehrani & Khorasani, 2009, 12–13; Patan 2008, 11–14)
The first step of physics-based modelling is always simplifying the considered process. This
is achieved with assumptions to simplify the calculations or to enable them at all with an
acceptable expenditure. After simplifying assumptions, the description of the system is for-
mulated in the form of either ordinary or partial differential equations, which can be classi-
fied as balance equations, physical or chemical equations of state, phenomenological equa-
tions or interconnection equations. Defining the system model often leads to complex equa-
tions, which can be simplified with mathematical treatments like linearization, approxima-
tion with lumped parameters and order reductions, before solving the obtained equations
using numerical methods. (Leturiondo 2016, 17; Isermann 2006, 71)
Physics-based models stem from natural laws, and as such, they have several advantages
over other approaches. Physics-based models are general models and can produce a very
precise description of the studied system and phenomena. These models do not depend on
comparisons to experimental data, making them not bound to certain environmental or op-
erational conditions. This does not only make physics-based models applicable to different
situations but also allows simulating new insertions to the system and produce results for
analysis. (Leturiondo 2016, 17; Isermann 2006, 71; Stringer et al. 2012, 156)
The downside of using physics-based models is their complexity. Creating the model is al-
ways balancing between accuracy and complexity and the associated cost and time. Creating
an extensive and accurate physics-based model can be practically impossible or the model
35
can become excessively complex, and despite the rapid development of computational
power, the calculation process too demanding for practical purposes. Therefore, simplifica-
tion of the modeled system, the model itself and the solved equation set is required to obtain
a technically feasible model. (Jardine et al. 2006, 1494; Stanley 2010)
5.2 Data-driven modelling
Unlike in physics-based modelling, data-driven models are not based on natural laws or
physics of the system. Instead, data-driven models are constructed from measurements and
rely on monitored and historical data that are used to train the system behavior to the model.
Since data-driven models have no causal relationship between the model inputs and outputs,
they are also referred as black-box models, whose reliability must be questioned. On the
other hand, data-driven models do not require a priori knowledge of the underlying physics
of the system and can thus be implemented to complex systems, where physics-based ap-
proach is not feasible. (Leturiondo 2016, 18; Sutharssan et al. 2015, 2)
Data-driven models are based on the assumption that the statistical characteristics of the
system do not change until a fault occurs (Pecht & Jaai, 2010, 318). This makes data-driven
models faster to implement and computationally more efficient than physics-based models.
The main disadvantage of data-driven models is the absolute dependency on historical and
empirical data, making these models applicable only to operating conditions in which his-
torical data is available. Inappropriate dataset selection for the model construction can lead
to false alarms (type I error) or inability to detect process faults (type II error). For reliable
diagnosis and prognosis, also run-to-failure data is required, which are often not available
for a particular system or component. Run-to-failure data sets are also used to generate re-
spective threshold values for monitoring purposes. Even though these data sets can be diffi-
cult to obtain, for instance to critical and expensive components or to new systems without
any historical data, they can be achieved with several techniques. These techniques can be
e.g. hardware-in-the-loop simulations, where the hardware is put under simulated loads as
in the real application or accelerated life test, where the component is put under elevated
stress to fail more quickly. (Ge 2017, 18; Sutharssan et al. 2015, 2)
36
Data-driven approaches can be classified as statistical approach or computational intelli-
gence (CI) approach. Section 5.2.1 describes the characteristics of statistical models and
some standard methods. Computational intelligence is a branch of artificial intelligence (AI)
widely used in different fields of research and engineering practice. Computational intelli-
gence uses computational methods inspired by processed found in nature and in the human
brain being one of the most effective data processing tools to solve complex nonlinear prob-
lems. Two of the most important CI methods used in condition monitoring applications are
artificial neural networks, which are discussed in Section 5.2.2, and approaches based on
fuzzy logic. (Zhang et al. 2017, 783; Das et al. 2012, 728–730)
A fuzzy logic system (FLS) is generally a nonlinear mapping of an input data vector into a
scalar output (Mendel 1995, 346). The process consists of four components presented in
Figure 14: fuzzifier, rulebase, interference engine, and defuzzifier. The first component is
used to fuzzify crisp input data to apply linguistic statements based on expert knowledge in
the form of IF-THEN rules. These rules are combined and applied in the interference engine,
which mimics the human decision making based on fuzzy concepts. The mapped fuzzy out-
put sets are finally defuzzified to produce a crisp output. (Song & Johns, 1997, 220–222)
Figure 14. Block diagram of a FLS (Largueh et al. 2015).
37
Fuzzy logic systems can effectively use inexact, subjective and ambiguous data and vague
knowledge elements giving the ability to map realistic situations. FLS imitates logical hu-
man thinking, making FLS applicable for qualitative analysis of complex large-scale systems
and is therefore used widely in many different industrial fields. In condition monitoring, FLS
are often used in conjunction with other methods to aid in the decision making process,
mainly due to slow interference speed and low precision. (Zhang et al. 2017, 784–785, 787–
788; Suganthi et al. 2015, 586–587; Shanmuganathan 2016, 12)
5.2.1 Statistical models
Statistical approaches to condition monitoring and fault detection are the oldest and most
widely employed strategies partly due to their easy development and implementation. Most
statistical approaches are built from normal operational data to represent normal healthy op-
eration. Analysis of the underlying statistical properties, such as variance and standard de-
viation of the monitored data, is carried out to detect changes in these parameters. Statistical
approaches can be divided into univariate approach and into multivariate approach. Tech-
niques under the former category monitor each characteristic independently, which makes
them impractical to deploy in modern industrial applications. Multivariate approaches do not
face the same challenge, since they monitor simultaneously many characteristics and also
take into account the interdependencies existing among these characteristics. (Das et al.
2012, 725; Sutharssan et al. 2015, 3; Severson et al. 2016, 191; Qin 2012, 221)
Multivariate techniques are powerful tools whose primary function is to convert a number
of related process variables to a smaller set of uncorrelated values. Multivariate techniques
have the ability to compress data and reduce its dimensionality while preserving essential
information making the analysis easier compared to the original larger data set. They are
also capable of handling noisy data and correlation to effectively extract true information.
Many techniques have been researched and implemented in numerous areas and industrial
fields, two of the most used techniques being principal component analysis (PCA) and partial
least squares (PLS). (Venkatasubramanian et al. 2003b ,331)
Principal component analysis is a dimensional reduction technique based on orthogonal de-
composition of the covariance matrix of the process variables along directions that explain
the maximum variation of the data. Dimensionality is lowered by finding factors having
38
lower dimension than the original data set and the ability to properly describe the major
trends in the original data set. (Venkatasubramanian et al. 2003b ,331)
The PCA model can be illustrated as presented by Sheriff et al. (2017, 239–241) and shown
in Figure 15, where 𝐗 denotes input data matrix with dimensions of 𝑛 x 𝑚, 𝑛 being the num-
ber of process variables and 𝑚 the number of observations. Using single value decomposi-
tion, the input data matrix can be expressed as
𝐗 = 𝐓𝐏T (2)
where 𝐓 is a matrix of the transformed variables and 𝐏 is a matrix of the orthogonal vectors.
The columns of the matrix 𝐏 are eigenvectors, also known as loading vectors, associated
with the covariance matrix of the input data matrix 𝐗, which is given by
∑=
1
𝑛 − 1𝐗T𝐗 = 𝐏𝐋𝐏𝐓 with 𝐏𝐏T = 𝐏T𝐏 = 𝐈𝑚
(3)
where 𝐋 is a diagonal matrix related to the 𝑚 principal components and 𝐈𝑚 is the identity
matrix. The number of principal components of the PCA model is the same as the number
of original process variables in the input data matrix, but since most industrial process vari-
ables are highly correlated, the process variations can be captured with fewer principal com-
ponents. The PCA model quality is determined by the number of obtained principal compo-
nents, too many leading to noise and too few to decreased prediction ability. Selection of the
number of principal components is therefore crucial in the model creation. Various tech-
niques, e.g. cumulative percent variance and cross validation, can be used to select appro-
priate number of principal components. In practice, usually the first two or three principal
components are enough to capture the variability and thus, the dimensionality is significantly
reduced. (Sheriff et al. 2017, 239–241; Venkatasubramanian et al. 2003b, 331)
Finally, the input data matrix can be expressed as
𝐗 = �̂��̂�T + �̃��̃�T = 𝐗�̂��̂�T⏞ �̂�
+ 𝐗(𝐈𝑚 − �̂��̂�T)⏞
𝐄
(4)
39
where �̂� ∈ R𝑛 x 𝑙 denotes the matrix of the 𝑙 retained principal components and �̃� ∈ R𝑛 x 𝑚−𝑙
the matrix of the ignored (𝑚 − 𝑙) principal components. Similarly, �̂� ∈ R𝑚 x 𝑙 and �̃� ∈
R𝑚 x 𝑚−𝑙 denote the matrices containing the 𝑙 retained eigenvectors and the ignored (𝑚 − 𝑙)
eigenvectors, respectively. Equation 4 expresses the original input data matrix with the mod-
eled variation �̂� using only the 𝑙 retained principal components and with matrix 𝐄 represent-
ing residuals formed by variations corresponding to process noise. (Sheriff et al. 2017, 240–
241; Mansouri et al. 2016, 2)
Figure 15. Schematic illustration of PCA (Sheriff et al. 2017, 241).
Partial least squares is another standard linear dimensionality reduction technique. Similarly
to PCA, PLS can be used with ill-conditioned data. The main advantage of PLS is that it
considers two sets of variables: predictor variables 𝐗 and response variables 𝐘. The former
represent process characteristics while the latter represent quality characteristics. PLS can
evaluate the relationship between 𝐗 and 𝐘 maximizing the covariance between them in the
reduced space defined by the number of latent variables and can be used as a regression tool
to predict the quality characteristics from the corresponding process characteristics. PLS
does not have a closed-form solution and thus uses an iterative algorithm. (Das et al. 2012,
726–727; Severson et al. 2016, 192; Tidriri et al. 2016, 66)
40
One standard algorithm, nonlinear iterative partial least square (NIPALS), can be summa-
rized as presented by Sheriff et al. (2017, 241–243), as follows. Score matrices and loading
vectors are calculated as:
𝐗 = 𝐒𝐏T + 𝐄 =∑𝒔𝑖𝒑𝑖T + 𝐄
𝑀
𝑖=1
(5)
𝐘 = 𝐔𝐐T + 𝐅 =∑𝒖𝑗𝒒𝑗T + 𝐅
𝑀
𝑖=1
(6)
where 𝐒 ∈ R𝑛 x 𝑀 is the orthonormal input score matrix and 𝐔 ∈ R𝑛 x 𝑀 the orthonormal out-
put score matrix; 𝐄 ∈ R𝑛 x 𝑚 and 𝐅 ∈ R𝑛 x p are the model residues; 𝐏 and 𝐐 are the loading
vectors of input matrix 𝐗 and output matrix 𝐘, respectively; 𝑚 and 𝑛 are the number of
process variables and observations in 𝐗; 𝑝 is the number of quality variables in 𝐘 and 𝑀 the
total number of extracted latent variables; 𝒔 and 𝒖 are the input and output score vectors;
and 𝒑 and 𝒒 are the input and output loading vectors.
First, the input and output matrices are standardized. To initialize the algorithm, one column
of the output data matrix is assigned as output score vector. At each iteration 𝒔, 𝒖, 𝒑 and 𝒒
are calculated and stored while 𝑀 latent variables are extracted. Iteration is continued until
convergence is reached. (Ibid)
The main limitations of both standard PCA and PLS techniques is their inability to cope with
nonlinear and dynamic processes. Numerous variants of both techniques suitable for nonlin-
ear and dynamic problems can be found in the literature, such as dynamic PCA and PLS,
dynamic nonlinear PCA and PLS, dynamic kernel PCA, and multiblock PCA (Das et al.
2012, 728; Tidriri et al. 2016, 65–66). PCA and PLS techniques are used together with con-
trol charts to assign appropriate threshold limits for alarms and detecting faults. The two
most widely used fault detection indices are the 𝑇2 and 𝑄 statistics. The 𝑇2 chart represents
the captured variation within the model plane indicating the distance of the current operation
from the desired operation. The desired state of operation is described by the model and the
corresponding principal components or latent variables. The 𝑄 statistics, also known as
squared prediction error (SPE), represents the variance not captured by the model. These
41
two charts must be used as a pair, as a significant deviation captured by either of the charts
indicates abnormal operation. The main advantage of using PCA and PLS together with the
control charts is that instead of monitoring many parameters at once, only the evaluation of
the two chart output variables is required to process monitoring (Yang 2004, 62). (Mac-
Gregor & Cinar, 2012, 112–113)
For fault diagnosis, PCA and PLS can also be used with the 𝑇2 and 𝑄 statistics, along with
their variants (Qiu & Bai, 2012). This approach usually has one PCA model for each system
behavior, i.e. for normal operation and for each faulty situation, which enables classification
of the model outputs (Villegas et al. 2010; Yang 2004, 67–69). Another practice for fault
diagnosis is to use the outputs of PCA and PLS models as inputs for other classification
tools, such as hidden Markov models (Zhou et al. 2004) or neural networks (Zhang et al.
2013). Other standard methods for diagnosis are contribution plots, which show the contri-
bution of each process variable to the calculated statistic (Westerhuis & et al. 2000, 95), and
reconstruction-based contributions (Fuentes-Garcia et al. 2018, 196).
5.2.2 Artificial neural networks
Artificial neural networks (ANN) are a widely used method, also in condition monitoring,
inspired by the human brain structure. They can represent highly nonlinear functions and
perform multi-input, multi-output mapping, making them very powerful in classification and
pattern recognition (Van Tung & Yang, 2009, 63). ANNs have two important features that
make them very valuable in many applications. First, ANNs can approximate any nonlinear
function with arbitrary accuracy with suitable architecture and weight parameters. Second,
they are capable of self-learning by extracting features from training data without a priori
knowledge about the studied process. (Patan 2008, 15)
The computational model of an ANN consists of processing elements called neurons, and
connections with weight parameter between neurons called links. Each neuron in the net-
work receives stimulus from the neighboring connected neurons, processes the stimulus and
creates output. A schematic ANN structure is illustrated in Figure 16, where the neurons are
arranged in different layers, namely input layer, one or multiple hidden layers and output
layer. The network can be structured, and the neurons can process information in various
42
ways, making ANN methods very versatile. (Ismail et al. 2016, 1; Zhang & Gupta, 2000,
61)
Figure 16. Simple feedforward ANN structure with two hidden layers (Ismail et al. 2016, 2).
ANN is trained with a training data set. During this process, the ANN learns the unknown
function and optimizes the network parameters by adjusting the weights according to input
and output observations (Van Tung & Yang, 2009, 63; Sutharssan et al. 2015, 5). The train-
ing can be supervised or unsupervised. In supervised learning, external input is required
about the desired output. For instance, common practice is to use data sets with known faults
to train an ANN to detect and diagnose these fault symptoms in actual operation. In unsu-
pervised learning such a priori knowledge is not required but the ANN learns itself using the
new available information. Another way to classify ANNs is based on the network architec-
ture, such as multilayer perceptrons (MLP), self-organizing maps (SOM) and radial basis
function networks, among others. The main limitations of any ANN are mostly due to weak
selection of network weights and parameters or the number of hidden neurons (Talebi et al.
2010, 8). (Khoukhi & Khalid, 2015, 19; Zhang & Gupta, 2000, 62)
43
MLP with back propagation training algorithm is one of the most used ANN approaches.
MLP has a feed-forward structure, as illustrated in Figure 16. Runtime data is fed to the input
neurons where from it propagates via connections to the hidden neurons in one or multiple
hidden layers and finally to the output neurons. Each neuron takes a weighted sum of its
inputs and the sum is fed to the neuron activation function to compute the final output of the
neuron, which can be the input of neurons in the following layer. The neuron activation
function can have several forms, such as linear thresholds and Gaussian functions, sigmoidal
functions, see Figure 17, being the most used one. (Van Tung & Yang, 2009, 63; Stanley
2013; Talebi et al. 2010, 7–11)
Figure 17. Common activation functions of MLP networks (Talebi et al. 2010, 9)
The nonlinear computing of an MLP illustrated in Figure 18 can be expressed as
𝒚 = 𝜎3{𝐖3𝜎2[𝐖2𝜎1(𝐖1𝒖)]}, (7)
where 𝜎1, 𝜎2 and 𝜎3 are vector valued activation functions; 𝐖1, 𝐖2 and 𝐖3 are the weight
coefficient matrices; 𝒖 is the input vector and 𝒚 the output vector (Patan 2008, 17).
44
Figure 18. A three-layer perceptron with 𝑛 inputs and 𝑚 outputs (Patan 2008, 18).
Training of an ANN is basically a process to determinate the weight coefficients that define
connection intensity between neurons. The fundamental back propagation training algorithm
is a form of supervised learning. The iterative algorithm uses the optimization gradient de-
scent method to minimize a sum-squared error. The weights inside the network are adjusted
based on the equation
𝒘(𝑘 + 1) = 𝒘(𝑘) − 𝛼∇𝐽(𝒘(𝑘)), (8)
where 𝒘(𝑘) denotes the weight vector at discrete time 𝑘, 𝛼 is the user-defined learning rate
and ∇𝐽(𝒘(𝑘)) is the gradient of the performance index 𝐽 with respect to the weight vector
𝒘. In the training phase, the error of the output layer is propagated towards the input layer
and the weights inside the network are adjusted based on the gradient of the performance
index. (Patan 2008, 18)
It is worthwhile to distinguish between shallow (one hidden layer) and deep (two or multiple
hidden layers) networks. While it is proven that a multilayer feedforward network with only
one hidden layer can approximate any nonlinear function (Hornik 1990) this can still be very
difficult or impossible in practice, since the required number of hidden neurons can grow
45
unfeasibly large. Deep neural networks can reduce the required number of hidden neurons
and they have become increasingly popular in recent years and proven very efficient in many
fields. (Goodfellow et al. 2016, 195–197).
Feedforward networks can functionally map input and output data points, but the standard
network is static (Malhi & Gao, 2004, 2048). Most of industrial processes are dynamic and
to take this into account, the ANN requires either external memory or feedback. The most
used network type with external memory is a multilayer perceptron type time-delay network,
which has two parts: a non-linear static approximator and an external dynamic filter bank.
These types of networks suffer from stability issues and are not as popular as recurrent net-
works with feedback loops. Recurrent neural networks are specialized for processing se-
quential data, which makes them convenient for time-series prediction, such as RUL predic-
tion. Recurrent networks can store information from previous time-steps by having feedback
connections from the hidden or output layers to the previous layers (Malhi & Gao, 2004,
2049). This network structure, illustrated in Figure 19, enables neuron to receive external
inputs and also state feedback signals from itself and other neurons in the network. Based on
the location of feedback loops, recursive networks can be divided into globally (feedback on
global network level) and locally (feedback on neuron level) recurrent networks. A variation
of the standard back-propagation training algorithm, called back-propagation through time
is usually applied to recurrent ANNs. (Patan 2008, 29–36; Goodfellow et al. 2016, 367–379)
46
Figure 19. A locally recurrent neural network. The hidden neurons receive external inputs (black arrows) and
feedback loops from previous time-steps (green arrows) creating output (red arrows). Adopted from Radha-
krishnan (2017).
For fault detection, ANNs can be used similarly to other methods by comparing measured
values to expected values. Usually the network is trained with healthy state data, and after
the training the network is run parallel to the monitored system to detect abnormalities in the
form of residuals. The residuals can also be used as input data for a second network for fault
isolation. The other used approach is based on a clear differentiation of various faults from
healthy state operation also including the differentiation between various faults. This ap-
proach uses supervised learning algorithms to identify a priori defined fault patterns in actual
operation. (Hush et al. 1997, 918–919; Das et al. 2012 ,729)
Fault diagnosis with ANNs is basically a classification problem. Instead of measurements or
process characteristics, fault patterns are used as network inputs producing classes or fault
categories. For example, the output classes can be labeled as healthy, anomaly and so on.
Regression can then be used to extrapolate the damage or fault propagation to estimate the
RUL of the monitored system. (Das et al. 2012 ,729; Sutharssan et al. 2015, 5)
47
5.3 Hybrid modelling
The above discussed modeling methods to condition monitoring each have their own ad-
vantages and disadvantages summarized in Table 2. Many different methods have been used
in conjunction to overcome these limitations. Combining different modeling methods is
known as hybrid modeling. Hybrid models are considered a promising approach, as different
models complement each other. Hybrid models can be constructed in various ways. For ex-
ample, neuro-fuzzy systems combining ANNs and fuzzy logic systems have been widely
used. This approach utilizes the learning capability of ANNs and the ability of fuzzy logic
systems to represent human knowledge (Van Tung & Yang, 2009, 62). Another approach is
to use physics-based models for fault detection combined with data-driven fault diagnosis
(Das et al. 2012, 737). Physics-based models can also be used to generate fault signals, which
can be used as training data for machine learning approaches. Such an approach was pro-
posed, for example, by Garcia-Matos et al. (2013), as they used a physical model to create a
fault dictionary to be compared with faults detected using an MLP model to monitor an axial
compressor. Other examples of hybrid models for power plants are proposed by Simani &
Fantucci (2000) and Chen et al. (2015). Simani & Fantucci proposed Kalman filters to re-
sidual generation and fault detection integrated with fault identification with ANNs and used
a physics-based power plant model for model testing. Chen et al. integrated generalized re-
gression neural network (GRNN), mean impact value (MIV), partial least squares regression
(PLSR) and B-spline transformation techniques to monitor sensors in power plants. The
GRNN model is used to assess the average contribution rate of independent variables and
the MIV method is used to filter out the main modeling parameters. These parameters are
then modeled using the PLSR method based on the B-spline transformation.
48
Table 2. Summary of advantages and disadvantages of different CM approaches.
Approach Advantages Disadvantages
Physics-based models can be highly accurate
explains process behavior
can be extrapolated to differ-
ent operation conditions
requires less data compared
to empirical models
complex to create and solve
simplifying assumptions re-
duce accuracy
requires determination of
various physics parameters
Statistical models no a prior process
knowledge is needed
easy to develop and imple-
ment
can provide confidence in-
tervals of new observations
resistant to input noise
often requires assumptions
of the underlying statistical
characteristics
cannot explain process be-
havior
requires large amount of
data
cannot be extrapolated to
different operation condi-
tions
Artificial neural networks no a prior process
knowledge is needed
can model complex nonlin-
ear processes
short development time
resistant to input noise
cannot explain process be-
havior
applicable only to operation
mode used for training
cannot detect and diagnose
noble faults
requires large amount of
data
training process is prone to
errors
6 CASE STUDY: DETECTION OF ABNORMALITIES WITH ARTI-
FICIAL NEURAL NETWORK BASED SOFTWARE
In this section, a commercial condition monitoring software Intelligent Health Monitor
(IHM), introduced in Section 6.1, is evaluated in practice in a case study with process data
from a reference power plant. The owner and location of the reference plant are confidential.
49
The gas-powered combined heat and power plant with gross thermal output of approximately
250 MW was commissioned in the 2000s.
6.1 Intelligent Health Monitor
Intelligent Health Monitor is a commercial condition monitoring software by algorithmica
technologies Gmbh (algorithmica technologies Gmbh 2018). IHM applies machine learning
in the form of recurrent neural networks to create a mathematical model for each parameter,
called tags, in terms of the others. The monitored parameters do not need to be actual meas-
urements. Tags can be any parameters included in the database, for example calculated pro-
cess parameters such as heat transfer coefficients derived from measurements. After the
model is created, dynamic alarm limits are employed to the monitored parameter. The dy-
namic alarm limits are computed from expected values and the chosen confidence interval
of the model. (Bangert 2018)
IHM is based on machine learning and requires training data from healthy state operation.
Ideally, the training data set would include all relevant operation conditions, i.e. load varia-
tions and startups, for the model to be accurate monitoring the actual operation. The inde-
pendent variables used to model each monitored parameter can be chosen manually by the
user or automatically based on a combination of correlation modeling and principal compo-
nent analysis. Once the model is created, it can be verified using second data set not included
in the training of the network. The model detects deviations from the expected healthy be-
havior and can be used in online condition monitoring or in offline analysis. (Ibid)
6.2 Case study: Decreased boiler efficiency
Boiler thermal efficiency is one of the key parameters to monitor in a thermal power plant.
The efficiency directly affects to the operating costs of the plant as boilers operating with
low efficiency require more fuel to generate the desired heat output. The boiler efficiency
can be calculated two ways. The direct way is to calculate the efficiency as the ratio of the
useful heat absorbed into the steam 𝑄abs and the heat and energy brought into the boiler 𝑄in
as:
50
𝜂 =
𝑄abs𝑄in
(9)
Another option is to calculate the efficiency indirectly as:
𝜂 = 1 −∑𝑄loss,𝑖 (10)
where 𝑄loss,𝑖 denotes individual heat losses. The hot flue gas out of the boiler represents the
largest individual heat loss, which is proportional to the mass flow and the temperature of
the flue gas. (Huhtinen et al. 1994, 92–99) In this case study, the flue gas temperature after
the final superheater before the economizer is monitored. The flue gas temperature before
the economizer has gradually increased, which indicates decreased performance of the boiler
heat surfaces prior to the economizer.
6.2.1 Problem description
Historical data of the flue gas temperature before economizer is monitored using 96 different
models. The used models are created using different parameters in the model creation. The
varied parameters are the training data period, the number of hidden layers and neurons in
the ANN architecture and the independent (input) variables used to model the monitored
parameter. The training parameters are shown in Table 3 and in Table 4. The used training
data periods are one month, four months and twelve months. The models are created using
two or four (software’s default setting) hidden layers. Number of hidden neurons in the first
hidden layer are chosen based on the number of independent variables. The chosen number
of hidden neurons in the first hidden layer are 0.8, 1.5, 2, and 3 times the number of inde-
pendent variables. The number of hidden neurons is halved in each consecutive layer. For
example, in the case of 30 hidden neurons in the first hidden layer, the second layer has 15
hidden neurons, the third layer has 8 hidden neurons, and the fourth layer has 4 hidden neu-
rons. Four different ways to choose the independent variables were chosen. 10 and 30 inde-
pendent variables are chosen automatically by the software based on PCA and correlation
analysis. In addition, 10 and 30 independent variables are chosen manually based on process
knowledge. The independent variables in each case are shown in Table 5. Some of the vari-
ables in Table 5 have parallel measurements, e.g. several redundant temperature measure-
ments at the same location. Also, similar components, such as superheaters, are mentioned
only once in the list, although several of them are included in the models.
51
Table 3. Training parameters for models with 2 hidden layers.
Training period: 1 month Training period: 4 months Training period: 12 months
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
1AA 10 Auto 8 1BA 10 Auto 8 1CA 10 Auto 8
1AB 10 Auto 15 1BB 10 Auto 15 1CB 10 Auto 15
1AC 10 Auto 20 1BC 10 Auto 20 1CC 10 Auto 20
1AD 10 Auto 30 1BD 10 Auto 30 1CD 10 Auto 30
1AE 30 Auto 24 1BE 30 Auto 24 1CE 30 Auto 24
1AF 30 Auto 45 1BF 30 Auto 45 1CF 30 Auto 45
1AG 30 Auto 60 1BG 30 Auto 60 1CG 30 Auto 60
1AH 30 Auto 90 1BH 30 Auto 90 1CH 30 Auto 90
1AI 10 Man 8 1BI 10 Man 8 1CI 10 Man 8
1AJ 10 Man 15 1BJ 10 Man 15 1CJ 10 Man 15
1AK 10 Man 20 1BK 10 Man 20 1CK 10 Man 20
1AL 10 Man 30 1BL 10 Man 30 1CL 10 Man 30
1AM 30 Man 24 1BM 30 Man 24 1CM 30 Man 24
1AN 30 Man 45 1BN 30 Man 45 1CN 30 Man 45
1AO 30 Man 60 1BO 30 Man 60 1CO 30 Man 60
1AP 30 Man 90 1BP 30 Man 90 1CP 30 Man 90
52
Table 4. Training parameters for models with 4 hidden layers.
Training period: 1 month Training period: 4 months Training period: 12 months
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
Model
Independ-
ent varia-
bles
Number of
hidden
neurons in
the first
hidden
layer
2AA 10 Auto 8 2BA 10 Auto 8 2CA 10 Auto 8
2AB 10 Auto 15 2BB 10 Auto 15 2CB 10 Auto 15
2AC 10 Auto 20 2BC 10 Auto 20 2CC 10 Auto 20
2AD 10 Auto 30 2BD 10 Auto 30 2CD 10 Auto 30
2AE 30 Auto 24 2BE 30 Auto 24 2CE 30 Auto 24
2AF 30 Auto 45 2BF 30 Auto 45 2CF 30 Auto 45
2AG 30 Auto 60 2BG 30 Auto 60 2CG 30 Auto 60
2AH 30 Auto 90 2BH 30 Auto 90 2CH 30 Auto 90
2AI 10 Man 8 2BI 10 Man 8 2CI 10 Man 8
2AJ 10 Man 15 2BJ 10 Man 15 2CJ 10 Man 15
2AK 10 Man 20 2BK 10 Man 20 2CK 10 Man 20
2AL 10 Man 30 2BL 10 Man 30 2CL 10 Man 30
2AM 30 Man 24 2BM 30 Man 24 2CM 30 Man 24
2AN 30 Man 45 2BN 30 Man 45 2CN 30 Man 45
2AO 30 Man 60 2BO 30 Man 60 2CO 30 Man 60
2AP 30 Man 90 2BP 30 Man 90 2CP 30 Man 90
53
Table 5. Independent variables in model creation. Numbers inside parentheses () refer to the total number of
measurements corresponding to similar components and/or parallel measurements.
10 Automatic 30 Automatic 10 Manual 30 Manual
Flue gas volumetric
flow
Boiler steam power
(2)
Combustion air volu-
metric flow
Feed water mass flow
Main steam mass flow
Fuel mass flow (2)
Fuel power
Boiler selection
switch
Flue gas volumetric
flow (2)
Boiler steam power
(2)
Combustion air volu-
metric flow
Feed water mass flow
(2)
Main steam mass flow
(2)
Fuel mass flow (3)
Fuel power
Boiler selection
switch
Flue gas mass flow
Combustion air mass
flow
Combustion air fan
current
Spray water mass flow
Steam temperature af-
ter superheater
Fuel gas strainer pres-
sure difference
NO2 mass emissions
(2)
Furnace pressure (3)
Flue gas temperature
to stack
Combustion air pres-
sure
Emission measure-
ment temperature
Superheater material
temperature
Drum level
Main steam pressure
Main steam tempera-
ture
Superheater material
temperature (3)
Steam temperature
after superheater (5)
Main steam pressure
Main steam tempera-
ture
Superheater material
temperature (9)
Steam temperature
after superheater (6)
Drum pressure (2)
Feed water pressure
Feed water tempera-
ture after economizer
Drum material tem-
perature (3)
Feed water tempera-
ture
Steam temperature
superheater drain (3)
Flue gas temperature
to stack (2)
54
6.2.2 Evaluation criteria
Each created model is evaluated based on graphical inspection of the model trend. The mod-
els are evaluated using the following hierarchical criteria:
1. Model fit to training data
2. Model fit to monitored healthy data
3. Abnormality detection
The model needs to meet all the above criteria to be considered reliable and accurate. The
evaluation criteria is hierarchical in a sense that each model has to meet the previous criteria
before it is reasonable to evaluate the next criterion. I.e., a model capable of detecting the
increased flue gas temperature is still useless in practice, if the model is not capable of rec-
ognizing healthy behavior during the training period or in monitoring phase.
Each of the three criteria are evaluated on the following scale:
1. Very poor
2. Poor
3. Fair
4. Good
5. Very Good
During the training period and healthy state monitoring phase the models are evaluated as
follows. Models with constant very significant deviation (over 10 Celsius degrees) from
measured values or with clearly inconsistent prediction are treated as very poor. Poor models
follow the measured trend to some extent but are characterized by constant significant devi-
ations (from 5 to 10 Celsius degrees) and occasional very significant deviations from the
measured values. Models with decent overall production are treated as fair. These models
can capture the process dynamics but are characterized by constant minor deviations (from
3 to 5 Celsius degrees) and occasional significant or very significant deviations. Models that
most of the time have accurate prediction with often occurring minor deviations are treated
as good. Very good models follow the measured trend very well and have only few minor
deviations from measured values. During the last phase, i.e. abnormality detection, the above
criteria is used. In addition, the model’s performance is judged based on the level of the
55
predicted trend. Grade “good” requires trend prediction with similar behavior as the meas-
ured values with an indication of the increased temperature. The model evaluation criteria
for each phase are summarized in Table 6 and examples of each fit are illustrated in Figure
20.
Table 6. Model evaluation criteria.
Fit to training/moni-
tored data
Abnormality detection
Very poor Constant inconsistent
prediction
Constant very signifi-
cant deviation
Cannot be detected,
model is not reliable
Poor Constant significant de-
viation
Occasional very signif-
icant deviation
Cannot be detected,
model follows measured
trend at the same or at
higher level
Fair Constant minor devia-
tion
Occasional significant
or very significant devi-
ation
Minor indication, model
follows measured trend
Good Often occurring minor
deviation
Indication that results in a
more detailed analysis
Very good Occasional minor devi-
ations
Clear indication
56
a) Very poor fit (Model 1AN)
a) Poor fit (Model 1CC)
b) Fair fit (Model 1CD)
320
330
340
350
360
370
380
𝑇[°
C]
Observed Model
320
330
340
350
360
370
380
320 330 340 350 360 370 380
Mea
sure
d
Predicted
R2 = 0.199
320
330
340
350
360
370
380
𝑇[°
C]
Observed Model
320
330
340
350
360
370
380
320 330 340 350 360 370 380
Mea
sure
d
Predicted
R2 = 0.375
320
330
340
350
360
370
380
𝑇[°
C]
Observed Model
320
330
340
350
360
370
380
320 330 340 350 360 370 380
Mea
sure
d
Predicted
R2 = 0.435
57
c) Good fit (Model 1CE)
d) Very good fit (Model 2CH)
Figure 20. Examples of the scale used to evaluate the model fit. Each plot corresponds to a period of three
months during the monitoring phase.
6.2.3 Results
The models need to have the grade of “good” or “very good” in each evaluated category to
be considered useful and accurate enough to be used in actual plant monitoring. 13 of the 96
models met this criterion. Model 2CH with training period of 12 months, 30 automatically
chosen independent variables, 4 hidden layers, and 90 hidden neurons showed very good
performance and was chosen as a reference model, see Figures 21–24. The evaluation results
for models with 2 hidden layers are shown in Table 7 and for models with 4 hidden layers
in Table 8.
320
330
340
350
360
370
380
𝑇[°
C]
Observed Model
320
330
340
350
360
370
380
320 330 340 350 360 370 380
Mea
sure
d
Predicted
R2 = 0.803
320
330
340
350
360
370
380
𝑇[°
C]
Observed Model
320
330
340
350
360
370
380
320 330 340 350 360 370 380
Mea
sure
d
Predicted
R2 = 0.965
58
Figure 21. Measured and predicted flue gas temperature before economizer with reference model 2CH over
35 months. The dashed lines mark the three periods: training, healthy state operation and abnormality detec-
tion.
Figure 22. Measured and predicted flue gas temperature before economizer with reference model 2CH during
the training period of 12 months.
320
340
360
380
400
420
440𝑇
[°C
]Observed Model
Training Healthy state Abnormal
320
340
360
380
400
420
440
𝑇[°
C]
Observed Model
59
Figure 23. Measured and predicted flue gas temperature before economizer with the reference model 2CH
during the monitoring phase of 10 months (assumed healthy state).
Figure 24. Measured and predicted flue gas temperature before economizer with reference model 2CH during
abnormal operation of 13 months.
320
340
360
380
400
420
440
𝑇[°
C]
Observed Model
320
340
360
380
400
420
440
𝑇[°
C]
Observed Model
60
Table 7. Case study evaluation results of models with 2 hidden layers. Models with good overall performance
are highlighted.
Training period: 1 month Training period: 4 months Training period: 12 months
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Abnor-
mality
detec-
tion
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Abnor-
mality
detec-
tion
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Abnor-
mality
detec-
tion
1AA Very
good Fair Poor 1BA Good
Very
good Poor 1CA Fair Fair Poor
1AB Very
good Fair Poor 1BB Good
Very
good Poor 1CB Fair Fair Poor
1AC Very
good Fair Poor 1BC
Very
good
Very
good Poor 1CC Fair Fair Poor
1AD Very
good Fair Fair 1BD Good
Very
good Poor 1CD Poor Fair Poor
1AE Very
good Fair Poor 1BE
Very
good
Very
good
Very
good 1CE Fair Fair
Very
good
1AF Very
good Poor Poor 1BF
Very
good
Very
good
Very
good 1CF Fair Fair Poor
1AG Very
good Fair Poor 1BG
Very
good
Very
good
Very
good 1CG Fair Fair
Very
good
1AH Very
good Fair Poor 1BH
Very
good
Very
good Fair 1CH Fair Good Poor
1AI Very
good
Very
poor
Very
good 1BI Fair Fair Poor 1CI Poor Poor
Very
poor
1AJ Very
good Poor Good 1BJ Good Fair Poor 1CJ Poor Poor
Very
poor
1AK Very
good
Very
poor
Very
poor 1BK
Very
good Fair Poor 1CK Poor Poor Poor
1AL Very
good
Very
poor
Very
poor 1BL
Very
good Fair
Very
poor 1CL Poor Fair
Very
poor
1AM Very
good
Very
poor
Very
poor 1BM
Very
good Poor Fair 1CM Poor Fair Poor
1AN Very
good
Very
poor
Very
poor 1BN
Very
good Fair Fair 1CN Fair Poor Poor
1AO Very
good
Very
poor
Very
poor 1BO
Very
good Fair
Very
poor 1CO Fair Fair Fair
1AP Very
good
Very
poor
Very
poor 1BP
Very
good Poor
Very
poor 1CP Good Fair Fair
61
Table 8. Case study evaluation results of models with 4 hidden layers. Models with good overall performance
are highlighted.
Training period: 1 month Training period: 4 months Training period: 12 months
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Ab-
nor-
mality
detec-
tion
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Abnor-
mality
detec-
tion
Model
Fit to
train-
ing
data
Fit to
moni-
tored
data
Abnor-
mality
detec-
tion
2AA Fair Good Fair 2BA Very
good
Very
good Poor 2CA
Very
poor
Very
poor
Very
poor
2AB Fair Poor Poor 2BB Very
poor
Very
poor
Very
poor 2CB
Very
poor
Very
poor
Very
poor
2AC Very
good Poor Fair 2BC
Very
good Good Fair 2CC Fair Poor Poor
2AD Very
good Poor Poor 2BD
Very
good Good Poor 2CD Poor Fair Poor
2AE Good Fair Poor 2BE Very
good
Very
good
Very
good 2CE Good Good Poor
2AF Very
good Poor Poor 2BF
Very
good Good Good 2CF
Very
good
Very
good Good
2AG Very
good Fair Poor 2BG
Very
good
Very
good Good 2CG
Very
good
Very
good
Very
good
2AH Very
good Fair Poor 2BH
Very
good
Very
good
Very
good 2CH
Very
good
Very
good
Very
good
2AI Fair Poor Poor 2BI Very
poor
Very
poor Poor 2CI
Very
poor
Very
poor
Very
poor
2AJ Very
good Poor
Very
good 2BJ
Very
good Fair
Very
poor 2CJ Poor Poor Fair
2AK Fair Very
poor
Very
good 2BK
Very
good
Very
good Poor 2CK Poor Good Poor
2AL Very
good Poor
Very
good 2BL
Very
good Fair Poor 2CL Poor Poor Poor
2AM Very
good
Very
good
Very
good 2BM
Very
good Poor
Very
good 2CM Fair Fair Fair
2AN Very
good
Very
poor
Very
good 2BN
Very
good Good Poor 2CN Fair Fair Fair
2AO Very
good
Very
poor
Very
good 2BO
Very
good Fair Fair 2CO
Very
good
Very
good
Very
good
2AP Very
good
Very
poor
Very
good 2BP
Very
good Poor
Very
poor 2CP Good Good
Very
good
62
Models with a training period of one month show similar behavior regardless of other train-
ing parameters. These models are characterized by good fit to the training data but inaccurate
prediction in actual monitoring. In this case, the training period of one month is too short for
the network to learn the process behavior. As a result from the small training data and com-
plex network structure, these models most likely suffer from overfitting. I.e., the models
have “memorized” the process behavior rather than learned the underlying statistics. Model
2AM (4 hidden layers, 30 independent variables and 24 hidden neurons in the first hidden
layer) is the only exception with very good performance.
Models with a training period of four months show better prediction capability. These mod-
els also fit very well to the training data while their performance on additional data sets
varies. Models 1BE–1BG and 2BE–2BH with 30 hidden neurons in the first hidden layer
have very good overall performance. These models form the best performing group as they
account for 7/13 of the reliable models.
The performance of models with a training period of 12 months varies the most. Models
with two hidden layers fail to predict the flue gas temperature during all three phases. The
model accuracy clearly increases with the added two hidden layers. There is still a lot of
variation among these models, as some models are reliable while the majority fails to learn
the temperature behavior.
In general, the model performance significantly varies, no matter how the models are clas-
sified. However, some conclusions can be made based on the results. First, models with 10
automatically or manually chosen independent variables are all inaccurate. In this case, with
the chosen architectures, the selected 10 input variables are not enough to capture the process
dynamics. Second, the 30 automatically chosen input variables seem to be most suitable of
the chosen options, even though these variables include some variables without any physical
relation to the monitored parameter. Third, models with a training period of one month have
the worst accuracy but increasing the training period from four to twelve months does not
seem to increase the models’ predictive power. The impact of the number of hidden neurons
on the model performance considerably varies and no conclusions on this relation can be
made.
63
7 ARTIFICIAL NEURAL NETWORK BASED CONDITION MONI-
TORING
The implementation of ANN based monitoring into a condition monitoring system can be
divided into three parts: training phase, verification phase and monitoring phase. This section
discusses the practical aspects that need to be considered in the training and model verifica-
tion phases. The monitoring phase is relatively straight-forward from technical point of view
and is not discussed in this section. The considered aspects are illustrated with the case study
from Section 6.
7.1 ANN training
The performance and accuracy of any ANN is determined by the quality of the network
training. ANNs are often described as “black-box” models, since the network’s data pro-
cessing is complex and often difficult to comprehend (May et al. 21, 2011). Commercial
software can strengthen this image as the user has limited possibilities to vary the training
parameters or the network structure and to investigate the created models. Nevertheless, it is
vital for the user to have basic understanding on how different parameters affect on the net-
work performance. It is also important to distinguish between different ANN types and spe-
cific engineering problems. This section considers ANN’s used in regression problems and
more specifically time-series prediction.
The training parameters of ANN’s are highly problem specific and no universal solution for
optimal parameter selection exists (Matignon 2005, 147). Desired training parameters result
in a model that is computationally efficient and generalizes well. Generalization is the most
desired network property, which means that the network can identify the dependencies of
the training data in new data sets introduced in the monitoring phase. Insufficient training
parameters may result in overtraining or overfitting. Overtraining refers to a situation, where
the network training is continued beyond optimal point (Tzafestas et al. 1996, 508) and over-
fitting to exceeding some optimal ANN size. In both situations, the network “memorizes”
the patterns, including noise, present in the training data set and loses the prediction capa-
bility (Tetko et al. 1995, 827). Overfitting results from using too complex network structure
with excessive degrees of freedom (Alman & Ningfang, 2001, 121). This problem stems
from the so-called curse of dimensionality meaning that if the dimension of input space is
64
high, an exponentially large number of observations is required for optimal generalization
and the network uses most of its recourses to represent irrelevant portions of the input space
(Bach, 2014, 2). It is therefore important to adjust the training parameters accordingly to the
monitored system. Both over-simplified and too complex models are to be avoided.
7.1.1 Data requirements and training period
All machine learning algorithms, including ANNs, learn the system behavior from training
data, which makes the selection of appropriate training data a vital part of the training pro-
cess. The first consideration is to determine what kind of operation conditions are monitored.
These same conditions, e.g. load and environmental conditions, should be present in the
training data. The second important consideration is to determine healthy state of the moni-
tored parameter. If the training data includes faulty behavior of the monitored equipment,
the model can learn the behavior and won’t recognize it as abnormal in the monitoring phase.
Careful inspection of the training data is therefore required to avoid training unwanted pro-
cess behavior to the model. This task can be challenging, especially for slowly developing
faults. The longer the available historical time period, the easier this task is.
The determination of healthy state is also closely related to the question of when and how
often the model should be updated. Most commercial condition monitoring software based
on machine learning can be categorized into continuously updated models and periodically
updated models. Software in the first category continuously adjust the network based on
constant acquisition of process data. Online data is fed to the system leading to continuous
training and update of the model. The advantage of this approach is that no manual work is
required in the update process. The major disadvantage is that the network may be over-
adjusted and as a result, the model will learn the faulty behavior as normal operation and
lose fault detection capability. Periodically updated models require a lot of manual work
from experienced personnel. This work includes the training procedure, which depending on
the software, can be very time-consuming. More importantly, the training period and corre-
sponding training data should be chosen with special care, which requires detailed
knowledge of the process and plant events and operation. This in turn requires manual in-
spection of the data and the resulting personnel cost can be high. In general, the training
frequency of periodically updated models and the corresponding personnel costs can have a
large proportion of the system upkeep costs.
65
Component’s failure rate usually follows the so-called bathtub curve, see Figure 25, where
the breakdown probability of a component is high in the beginning of the plant operation
due to design defects and installation errors. The failure rate decreases during the early life
period, which is followed by the useful life period when the failure rate remains constant
until equipment start to suffer from mechanical stress and wear and the failure rate increases.
(Marlow et al. 2017, 716) Often a good starting point to network training is to use data from
the early stages of normal operation when process equipment are new and are considered
healthy. Another possibility is to update the model after revisions when the process equip-
ment are maintained. It is also clear that after process changes the used model should be
retrained. The obvious challenge in these situations is the lack of historical data. Physical
process simulators could be used to obtain the required training data but in practice, such
excessive and accurate simulators are hardly available. This means that often after process
changes the only option is to use the existing ANN model until new training data is available.
This needs to be considered when developing a condition monitoring system and for these
situations other monitoring methods should be reserved.
Figure 25. Theoretical bathtub curve (Singh & Adachi, 2013, 107).
The third practical question in model training is the length of the training period and the
corresponding size of the training data set. The availability of historical data can set practical
limitations to the training period but in general, as illustrated in the case study in Section 6,
longer training periods result in better models. Short training periods have the fundamental
66
issue that they rarely represent all the relevant operating conditions and process behaviors
that should be monitored in a complex power plant environment. Short training periods can
also lack the required number of samples required for the model to learn the system behavior.
Combination of a short training period and a complex model structure is also prone to over-
fitting. Increasing the training period too much has the drawback that the training may not
converge optimally. Increasing the training period also increases the required computational
effort, which however, rarely is a limiting factor in condition monitoring applications.
As with other ANN training parameters, the optimal training period length always depends
on the monitored system. Systems with simpler dynamics and fewer changing conditions
can be successfully modeled with shorter training periods than systems with more complex
behavior. The optimal training period also depends on the overall complexity of the network.
Models with high degree of freedom resulting from a large independent variable set or large
number of hidden neurons/layers should be trained with larger training data sets to avoid
overfitting. It is also worthwhile to note that the training period does not need to be contin-
uous. Instead, it is often advisable to select multiple training data sets corresponding to dif-
ferent conditions, e.g. seasons and loads.
The importance of adequate data preprocessing should not be underestimated. Usually the
preprocessing is integrated into commercial software packages or at least, the software de-
veloper provides instructions and requirements for the preprocessing.
7.1.2 Input variable selection
The second fundamental consideration in the ANN training is the selection of input varia-
bles, which has a major impact on the network’s performance and accuracy. Three aspects
make the input variable selection (IVS) challenging: The number of available variables is
usually very large in power processes; many variables have strong correlations with each
other; and many variables have weak predictive power. Ideally, the independent variables
are selected in a way that the set of variables has high predictive power, the variables are
dissimilar to each other with minimal redundancy and no irrelevant variables are present
increasing noise in the training. (May et al. 2011, 19)
67
IVS algorithms in the literature can be categorized in three main classes: wrapper, embedded
or filter algorithms (May et al. 2011, 23). Automatized IVS algorithms can be useful in some
instances and some algorithms are often included in commercial software packages. In con-
dition monitoring applications, the software user should have major a priori knowledge of
the monitored system and some estimation of suitable candidate input variables. This
knowledge should be the basis of variable selection, which can be augmented with IVS al-
gorithms. Relying purely on automatized algorithms is a tempting option, especially when a
large number of models are used. This can be successful, as in the case study in Section 6,
but in general, such algorithms should be used with caution. There is no guarantee of any
physical relation between automatically selected input variables and the predicted output
variable. A common problem with time series is that an input variable 𝑥1 correlating to the
output variable 𝑦 actually have a mutual dependency on a variable 𝑥2 (Fawcett, 2015), such
as time or power output. In such cases, variable 𝑥1 correlates with the output variable but
has little generalized predictive power and is usually weak to detect faulty situations.
The selection of independent variables is a challenging task even with sufficient knowledge
of the studied system. The variables may be selected in multiple ways depending on the
monitored system. One way to choose the independent variables is to introduce a distance
measure to each monitored parameter defining how distant measurements upstream and
downstream relative to the monitored parameter are included. Nearby measurements should
correlate well to the output parameter and reduce noise. The risk of using only measurements
very close to the output variable is that the model may not learn all the relevant process
dynamics. The choice of how distant measurements are included also depends on the moni-
toring strategy. If only a small number of models are used, more distant measurements
should be included in the models to detect faulty behavior. This is because nearby measure-
ments result in very specific models that are capable of detecting anomalies at a specific
location. If a fault has occurred further away from the monitored component the fault still
most likely remains unnoticed and another model at that specific location is required.
The number of independent variables has a significant impact on the network’s performance.
Ideally the independent variable set is kept as small as possible while still preserving predic-
tive power. This way the network’s complexity is not unnecessarily increased, and chance
68
of overfitting is decreased. Also, the iteration of the network’s weight coefficients is more
likely to converge, and computational effort is decreased. However, larger set of independent
variables usually provide more flexibility and possess more predictive power as more fea-
tures are included in the model, which is often useful for more complex systems. This is
visible in the case study results, as models with 30 independent variables clearly outperform
models with 10 independent variables.
In order to keep the number of independent variables as small as possible and to decrease
the chance of overfitting, redundant variables are to be avoided. Power plants often have
several parallel measurements of the same quantity, e.g. temperature or pressure measure-
ments at the same location. These parallel measurements represent the same feature and in-
cluding them all in the model only increases the model complexity while bringing little or
no additional improvement. It is therefore advisable to select only one of the parallel meas-
urements or use their average. Similarly, the set of independent variables should, as the name
states, be independent. If the input variables are linearly dependent, the same variable is
effectively introduced as multiple inputs. Commercial software often apply a sensitivity
analysis to the set of independent variables showing how significant impact each input var-
iable has on to the output variable. This analysis can be used to select the most appropriate
parallel measurement and to help choosing the candidate variables with highest predictive
power. However, the relation between the input and output variables is very complex to
measure and therefore, such algorithms should be treated with care.
7.1.3 Network architecture
The network architecture is defined by the number of hidden layers, hidden neurons in these
layers and the connections between the neurons. While it is shown that a multilayer feedfor-
ward network with only one hidden layer is capable of approximating any nonlinear function
(Hornik 1990), additional hidden layers may be helpful in the model training with complex
data sets, such as time-series (Heaton 2017).
Again, the optimal network architecture is problem specific and depends in a complex way
on the network type, the complexity of the learning task, neuron activation functions, train-
ing algorithm and other training parameters. Some rules-of-thumb for neuron selection have
been presented, such as the number of hidden neurons should be less than twice the size of
69
the input layer (Heaton 2017). Several statistic methods have also been created, some of
which are presented in Table 9. Even though such methods have been proposed, they are
difficult to implement when using commercial software packages. These methods are spe-
cific for certain criteria, require strict a priori assumptions and are still crude approximates
at best. Ultimately, finding a working network architecture comes down to trial and error.
Table 9. Various approaches to fix the number of hidden neurons. 𝑁h denotes the number of hidden neurons,
𝑛𝑖 the number of inputs, 𝑛𝑜 the number of outputs, 𝑁 the input-target relation, 𝑛𝑝 the number of input samples
and 𝐿 the number of hidden layer. Adopted from Sheela & Deepa (2013, 9).
Method Year Number of hidden neurons
Li et al. method 1995 𝑁ℎ = (√1 + 8𝑛𝑖 − 1) 2⁄
Tamura and Tateishi method 1997 𝑁ℎ = 𝑁 − 1
Zhang et al. method 2003 𝑁ℎ = 2𝑛𝑖 𝑛𝑖⁄ + 1
Jinchuan and Xinzhe method 2008 𝑁ℎ = (𝑛𝑖 +√𝑛𝑝)/𝐿
Shibata and Ikeda method 2009 𝑁ℎ = √𝑛𝑖𝑛𝑜
Sheela and Deepa method 2013 𝑁ℎ = (4𝑛𝑖2 + 3) (𝑛𝑖
2 − 8)⁄
Often the best network architecture is the simplest one. It is therefore advisable to start model
training with a simple architecture. Often a practical starting point is provided by the soft-
ware’s default settings or recommended architecture. If the created model is not satisfactory,
the architecture complexity can be progressively increased by adding hidden neurons and/or
layers.
7.2 Model verification
Validation and verification of used ANN models is important for a reliable CMS. Validation
refers to checking the model performance during training. Validation is dependent on the
used training algorithm and is usually automated in commercial software. Verification refers
to ensuring the model performance after training and is discussed in this section.
70
Ideally, every used model should be verified with two additional data sets: healthy state data
and faulty state data. Statistical measures can be used to evaluate the model’s prediction in
comparison to the healthy state data. Some of the most used measures include mean absolute
error (MAE), mean squared error (MSE) and root mean squared error (RMSE), whose results
are summed over the whole data set and normalized by the size of the data set. The metrics
are defined by the following equations:
𝑅𝑀𝑆𝐸 = √∑∑(𝑜𝑖𝑗 − 𝑧𝑖𝑗)2
𝑚
𝑗=1
𝑛
1=1
(11)
𝑀𝑆𝐸 =∑ ∑ (𝑜𝑖𝑗 − 𝑧𝑖𝑗)
2𝑚𝑗=1
𝑛𝑖=1
𝑛𝑝 (12)
𝑀𝐴𝐸 =
∑ ∑ |𝑜𝑖𝑗 − 𝑧𝑖𝑗|𝑚𝑗=1
𝑛𝑖=1
𝑛𝑝 (13)
where 𝑛𝑝 is the number of patterns in the data set, 𝑚 is the number of outputs, 𝑜 is the output
of a single neuron 𝑗, 𝑧 is the target for the single neuron 𝑗, and 𝑖 denotes input patterns
(Twomey & Smith, 1996). Another useful way to evaluate the model performance is to use
scatter plots and calculate the coefficient of determination 𝑅2 as
𝑆𝑆res =∑(𝑦𝑖 − 𝑓𝑖)2
𝑖
(14)
𝑆𝑆tot =∑(𝑦𝑖 − �̅�)2
𝑖
(15)
𝑅2 = 1 −
𝑆𝑆res𝑆𝑆tot
(16)
where 𝑦 denotes predicted values and 𝑓 measured values (Rasila 2007).
The 𝑅2 together with scatter plots provide a simple tool to evaluate the model prediction.
However, it should be highlighted, that 𝑅2 is a relative measure for the goodness of fit and
should not be used as only performance metric (Reisinger 1997, 1). Figure 26 shows the
reference model 2CH prediction vs. measured values during each phase. The high 𝑅2 values
indicate very high prediction accuracy. The plots are also useful in determining weak areas
71
in the model performance. The highest deviation between measured and predicted values
occur with low flue gas temperatures. In this case, these values are not interesting, since they
correspond to plant downtime, but similar situations may be interesting in other applications.
The effect of plant downtime and other similar situations that are not relevant from the per-
spective of condition monitoring should be taken into account when evaluating the model
accuracy. These situations are often not included in the model training. If data periods cor-
responding to these situations are present in statistical model evaluation, they can skew the
performance metrics.
a) All data
b) Training period
c) Healthy state monitoring
d) Abnormality detection
Figure 26. Measured vs. predicted values of the reference model 2CH with very good performance.
Verification of healthy state prediction is relatively simple if only sufficient amount of data
is available. Commercial software packages usually include some statistical measures to
0
100
200
300
400
500
0 100 200 300 400 500
Mea
sure
d
Predicted
R2 = 0.986
0
100
200
300
400
500
0 100 200 300 400 500
Mea
sure
d
Predicted
R2 = 0.999
0
100
200
300
400
500
0 100 200 300 400 500
Mea
sure
d
Predicted
R2 = 0.997
0
100
200
300
400
500
0 100 200 300 400 500
Mea
sure
d
Predicted
R2 = 0.989
72
evaluate the model performance and the above described performance metrics are also sim-
ple to implement using standard tools, such as Microsoft Excel. Another way is to use graph-
ical inspection and linguistic statements, as in the case study in Section 6. This approach is
not as objective or precise as statistical measures but is still sufficient enough in many cases.
Linguistic statements based on graphical inspection provide often the simplest way to model
evaluation, which is desirable when creating a large number of models. The model training
typically requires creating many models with different parameters and many personnel may
be involved in this process. Linguistic statements also provide the easiest way to share in-
formation between involved personnel without the need to interpret any statistical metrics.
The main challenges in the model verification arise from defining the required model per-
formance level and from the lack of faulty state data. Regardless of the performance metrics
used, the final decision of defining which level of model accuracy is required, is determined
by the software user. The goal of using ANN models in plant-wide condition monitoring is
to detect abnormalities during operation and also this capability should be verified. The case
study results underline this requirement, as several models had accurate prediction during
the training and healthy state operation but were unable to detect the increased flue gas tem-
perature. Verifying this capability can be challenging, since the available data usually lacks
clear failures. If a large number of components is monitored, it is practically impossible to
verify the fault detection capability of each created model. However, it is highly beneficial
to verify this capability with as many components as possible to establish trust in the moni-
toring approach.
7.3 Model training and verification summary
Finding optimal network structure, independent variables and training parameters is a diffi-
cult task that depends on many factors in a complex way. No universal solution exists and
the procedure involves many trials and errors. In plant-wide condition monitoring, the actual
task is often not finding an optimal ANN model. In most cases, it is enough to find a model
accurate enough to detect abnormalities without excessive false alarms during healthy state
operation for the operators to find the model reliable. To establish this trust, each used model
should be verified with healthy state data and at least one model with faulty state data.
73
It is vital to understand that the training parameters depend on each other, e.g. using a large
number of hidden neurons typically requires large training data set. It is also worth to notice
that a reliable model can be trained with several different parameter combinations, but on
the other hand, there is no guarantee the same parameters result in an accurate model on
some other component. In addition, information flow of detailed plant events should be pro-
vided to the software user to assist in the training period selection to avoid introducing faulty
state operation in the training phase.
As the training process is iterative by nature and the aim is to monitor a wide range of equip-
ment, it is important to create general guidelines for the model training with a specific soft-
ware. Initial systematic in-house testing with different network structures and training pa-
rameters should be carried out to determine a practical starting point defining initial training
parameters for the model training. These initial training parameters can be used in the pro-
posed general flowchart presented in Figure 27. Documented testing and acquired experi-
ence will significantly reduce the workload when monitoring is extended to additional plants.
74
Figure 27. General flowchart for ANN model training.
8 CONCLUSIONS
Power plant maintenance has been traditionally organized according to corrective and pre-
ventive maintenance strategies, which both have their limitations that can lead to high costs
due to low availability and inefficiently timed maintenance actions. CBM is considered an
advanced maintenance strategy and is already applied in many critical components in mod-
ern power plants. Extending condition monitoring and more efficient maintenance practices
to a wider range of equipment is desirable but requires advanced condition monitoring meth-
ods that can detect changes in process parameters indicating faults and malfunctions. These
methods include different mathematical models that can be classified as physics-based, data-
driven and hybrid models.
75
Artificial neural networks are a popular data-driven approach due to their self-learning ca-
pability and their ability to approximate nonlinear functions. ANN’s can be used in many
ways in condition monitoring applications including pattern recognition, classification and
time series prediction. Many commercial condition monitoring software based on ANN’s
are available for asset owners and O&M service providers. In this thesis, a commercial soft-
ware Intelligent Health Monitor (algorithmica technologies GmbH) was evaluated in a case
study to monitor a boiler’s increased flue gas temperature, which indicates decreased per-
formance of heat surfaces prior to economizer. The software is based on recursive neural
networks that are trained with historical data from healthy state operation.
A total of 96 models were created to monitor historical data of the flue gas temperature. Each
model was trained using different parameters and evaluated based on graphical inspection
of the model trend. The varied parameters were the length of the training data period, the
network architecture including the number of hidden neurons and layers, and the independ-
ent variables used to predict the output variable. The historical data of three years was di-
vided into three periods: training period, healthy state operation and faulty state operation
when the flue gas temperature had gradually increased. 13 of the 96 models were considered
reliable, as they were able to accurately predict the healthy state operation during training
and monitoring phases and also detect the increased flue gas temperature. In general, the
results of the case study significantly varied but two conclusions based on the results can be
made: training period of one month is too short for a reliable model and 30 automatically
chosen independent variables based on PCA and correlation analysis led to best performing
models.
The case study showed that ANN based software can be successfully used to detect abnor-
malities in power plant environment. However, the case study only includes the monitoring
of a single parameter and the applicability to other monitored parameters should be studied
further. The case study also showed that the network training and verification are case-spe-
cific and can be very challenging. The following aspects should be considered in the model
training and implementation to a condition monitoring system.
76
1. No universal solution for optimal network structure or parameter selection exists.
The training parameters depend on many factors in a complex way, e.g. system com-
plexity, network complexity, available training samples, and used training algorithm
and the training parameters also affect each other. In general, lots of trials and errors
are to be expected in the model training.
2. Systematic in-house testing is advisable to determine practical initial training param-
eters for a specific software. Power plants have a range of similar components and
often asset owners and O&M service providers monitor several plants. Well-docu-
mented testing can provide a practical starting point for model training, which can
significantly reduce the work load when extending condition monitoring to addi-
tional plants.
3. Each used model should be verified with additional data set corresponding to healthy
state operation. In addition, the fault detection capability should be verified, since
accurate healthy state prediction does not guarantee that the model can recognize
abnormal situations. This verification can be challenging, since such data sets are
often difficult to obtain.
The research objective of this thesis was met. The theoretical part gives a broad overview of
power plant maintenance and related modelling methods providing a fundamental introduc-
tion to the topic. The case study results and presented conclusions give valuable information
for O&M service providers and asset owners seeking new approaches for more advanced
condition monitoring. While ANN based condition monitoring is an attractive option for
plant-wide monitoring, many stakeholders hesitate to implement ANN based condition mon-
itoring into their condition monitoring systems due to the black-box nature of ANNs and due
to the fact that hardly any objective evaluation of commercial software are available. This
work bridges the gap between research and practice by offering practical considerations to
the implementation of ANN based condition monitoring and by providing documented soft-
ware evaluation.
77
8.1 Future work
Similar testing as in the case study should be carried out with other monitored parameters.
The case study showed that 30 automatically chosen independent variables, training period
of four months and hidden neurons arranged as 45-22-11-5 in four hidden layers resulted in
an accurate model. These same initial training parameters should be tested with other moni-
tored parameters to see if they can be used as a practical starting point in the model creation.
Adequate training and verification data are often difficult to obtain. One way to obtain such
data sets is to use physics-based process simulators and this option should be studied. An-
other practical research topic is the utilization of different modelling methods in condition
monitoring systems. O&M service providers require tools for different levels of condition
monitoring and the combined use of different software could provide flexible practical so-
lutions for condition monitoring in power plants.
REFERENCES
Akula Aparna, Goel Sudhanshu, Ghosh Ripul. 2017. Condition Monitoring Saves Money
and Prevents Failures. Chemical Engineering Progress. Vol. 113, Iss. 8, pp. 58–62.
Al-Turki Umar, Ayar Tahir, Yilbas Bekir Sami, Sahin Ahmet Ziyaettin. 2017. Integrated
Maintenance Planning in Manufacturing Systems. Cham: Springer. ISBN 978-3-319-06289-
1.
Algorithmica technologies GmbH. Intelligent Health Monitor. 2018. [Online]. [Accessed on
26.3.2018]. Available: http://www.algorithmica-technologies.com/en/solutions/intelligent-
health-monitor
Alman David H. & Ningfang Liao. 2001. Overtraining in Back-Propagation Neural Net-
works: A CRT Color Calibration Example. Color Research and Application, Vol. 27, No. 2,
pp. 122–125.
Bach Francis R. 2014. Breaking the Curse of Dimensionality with Convex Neural Networks.
eprint arXiv:1412.8690
Bangert Patrick. 2012. Optimization for Industrial Problems. Heidelberg: Springer. ISBN
978-3-642-24973-0.
Bangert Patrick. 2018. Smart Condition Monitoring Using Machine Learning. algorithmica
technologies GmbH. [Online]. [Accessed on 23.11.2017]. Available: http://www.algorith-
mica-technologies.com/system/case_studies/pdf_ens/000/000/022/original/16_Smart_Con-
dition_Monitoring.pdf?1498681861
Bania R.K. 2014. Survey on Feature Selection for Data Reduction. International Journal of
Computer Applications. Vol. 94, No. 18.
Bengtsson Markus, Olsson Erik, Funk Peter & Jackson Mats. 2004. Technical Design of
Condition Based Maintenance System – A Case Study using Sound Analysis and Case-
Based Reasoning. Proceedings of the 8th Conference of Maintenance and Reliability.
Bousdekis Alexandros, Magoutas Babis, Apostolou Dimitris & Mentzas Gregoris. A proac-
tive decision making framework for condition-based maintenance. Industrial Management
& Data Systems, Vol. 115 Iss. 7, pp. 1225–1250.
Chanda Pradip & Mukhopaddhyay Suparna. 2016. Operation and Maintenance of Thermal
Power Stations. India: Springer. ISBN 978-81-322-2720-5.
Chen Jianhong, Li Honkung, Sheng Deren & Li Wei. 2015. A hybrid data-driven modeling
method on sensor condition monitoring and fault diagnosis for power plants. International
Journal of Electrical Power & Energy Systems. Vol. 71, pp. 274–284.
Chookah M., Nuhi M. & Modarres M. A probabilistic physics-of-failure model for prognos-
tic health management of structures subject to pitting and corrosion-fatigue. Reliability En-
gineering & System Safety. Vol. 96, Iss. 12, pp. 1601–1610.
Crespo Adolfo Márquez. 2007. The Maintenance Management Framework. London:
Springer. ISBN 978-1-84628-820-3.
Das Anupam, Maiti J. & Banerjee R.N. 2012. Process monitoring and fault detection strate-
gies: a review. International Journal of Quality & Reliability Management, Vol. 29, Issue 7,
pp. 720–752.
Di Maio Francesco, Colli Davide, Zio Enrico, Tao Liu & Tong Jiejuan. 2015. A Multi-State
Physics Modeling approach for the reliability assessment of Nuclear Power Plants piping
systems. Annals of Nuclear Energy. Vol 80, pp. 151–165.
El-Thalji Idriss & Jantunen Erkki. 2015. A summary of fault modelling and predictive health
monitoring of rolling element bearings. Mechanical Systems and Signal Processing, Vol.
60–61, pp. 252–272.
EN 13306:2001. 2001. Maintenance terminology. European Standard. CEN (European
Committee for Standardization), Brussels.
Fawcett Tom. 2015. Avoiding Common Mistakes with Time Series. Silicon Valley Data
Science. [Online]. [Accessed on 18.3.2018]. Available: https://svds.com/avoiding-common-
mistakes-with-time-series/
Fedele Lorenzo. 2011. Methodologies and Techniques for Advanced Maintenance. London:
Springer. ISBN 978-0-85729-102-8.
Fortum. 2017a. Annual Review 2016. [Online]. [Accessed on 20.1.2018]. Available:
http://annualreport2016.fortum.com/en/
Fortum 2017b. Performance, OM and energy efficiency services for thermal power plants.
[Online]. [Accessed on 20.1.2018]. Available: https://www3.fortum.com/products-and-ser-
vices/power-plant-services/services-thermal-power-plants/performance-om-and-energy
Frisk Erik. 2001. Residual Generation for Fault Diagnosis. Doctoral thesis. Linköping Uni-
versity. Department of Electrical Engineering. ISBN 91-7373-115-3.
Fuentes-Garcia Marta, Macia-Fernandez Gabriel & Camacho Jose. 2018. Evaluation of di-
agnosis methods in PCA-based Multivariate Statistical Process Control. Chemometrics and
Intelligent Laboratory Systems. Vol. 172, pp. 194–210.
Garcia Salvador, Luengo Julian & Herrera Francisco. 2015. Data Preprocessing in Data Min-
ing. Cham: Springer. ISBN 978-3-319-10246-7.
Garcia-Matos Jesus, Sanz-Bobi Miguel, Munoz Antonio, Sola Antonio. 2013. Hybrid
Model-Based Fault Detection and Diagnosis for the Axial Flow Compressor of a Combined-
Cycle Power Plant. Journal of Engineering for Gas Turbines and Power. Transactions of the
ASME. Vol. 135, Iss. 5, pp. 1–5.
Ge Zhiqiang. 2017. Review on data-driven modeling and monitoring for plant-wide indus-
trial processes. Chemometrics and Intelligent Laboratory Systems 171, pp. 16–25.
Goofellow Ian, Bengio Yoshuaa and Courville Aaron. 2016. Deep Learning. MIT Press.
Gräber Ulrich. 2004. Advanced Maintenance Strategies for Power Plant Operators – Intro-
ducing Inter-plant Life Cycle Management. International Journal of Pressure Vessels and
Piping 81, pp. 861–8625
Guillén A.J. Cresp A. Macchi M. & Gómez J. 2016. On the role of Prognostics and Health
Management in Advanced Maintenance Systems. Production Planning & Control Vol. 27,
No. 12, pp. 991–1004.
Haddad Gilbert, Sandborn Peter & Pecht Michael. 2012. An Options Approach for Decision
Support of Systems With Prognostics Capabilities. IEEE Transactions on Reliability Vol.
61, No. 4, pp. 872–883.
Hashemian H.M. & Bean Wendell C. State-of-the-Art Predictive Maintenance Techniques,
in IEEE Transactions on Instrumentation and Measurement, vol. 60, no. 10, pp. 3480–3492.
Heaton Jeff. 2017. The Number of Hidden Layers. Heaton Research. [Online]. [Accessed on
12.3..2018]. Available at: http://www.heatonresearch.com/2017/06/01/hidden-layers.html
Hira Zena M. & Gillies Duncan F. 2015. A Review of Feature Selection and Feature Extrac-
tion Methods Applied on Microarray Data. Advances in Bioinformatics, 2015, 198363.
Holmberg Kenneth, Adgar Adam, Arnaiz Aitor, Jantunen Erkki, Mascolo Julien, Mekid
Samir. 2010. E-maintenance. London: Springer. ISBN 978-1-84996-205.
Hornik Kurt. 1990. Approximation Capabilities of Multilayer Feedforward Networks.
Neural networks, Vol. 4, pp. 251–257.
Huhtinen Markku, Kettunen Arto, Nurminen Pasi & Pakkanen Heikki. 1994. Höyrykattila-
tekniikka. Opetushallitus: Helsinki. ISBN 951-37-1327-X.
Hush D.R., Abdallah C.T., Heileman G.L. & Docampo D. 1997. Neural Networks in Fault
Detection: A Case Study. Proceedings of the American Control Conference. Vol 2, pp. 918–
921.
Isermann Rolf. 2006. Fault-Diagnosis Systems. Springer: Berlin. ISBN 978-3-540-24112-6.
Ismail Rahmat Izaibi B., Alnaimi Firas B. Ismail & AL-Qrimli Haidar F. 2016. Artificial
Intelligence Application in Power Generation Industry: Initial considerations. International
Conference on Advances in Renewable Energy and Technologies. IOP Conf. Series: Earth
and Environmental Science 32 (2016).
Jaako Juha. 1998. Aspects of process modelling. University of Oulu. Control Engineering
Laboratory. ISBN 951-42-5035-4.
Jardine Andrew, Lin Daming & Banjevic Dragan. 2006. A Review on Machinery Diagnos-
tics and Prognostics Implementing Condition-Based Maintenance. Mechanical Systems and
Signal Processing, Vol. 20, pp. 1483–1510.
Kaija Matti. 2016. Utilization of condition monitoring in power plant operation and mainte-
nance services. Master of Science Thesis. Tampere University of Technology.
Khoukhi Amir & Khalid Mohamed. 2015. Hybrid computing techniques for fault detection
and isolation, a review. Computers and Electrical Engineering. Vol 43, pp. 17–32.
Kobbacy Khairy & Murthy Prabhakar. 2008. Complex System Maintenance Handbook.
London: Springer. ISBN 978-1-84800-010-0.
Kylliäinen Ville-Veikko, Laaksonen Leila & Viitasaari Osmo. 2011. Käytettävyys, kunto ja
elinikä – kunnossapitostrategia tuottavuuden parantamisessa. Promaint. 5/2011, pp. 10–15.
Largueh Samia, Aloui Sinda, Pages Oliver & Chaari Abdessattar. 2015. Fuzzy Sliding Mode
Control For Turbocharger Diesel Engine. Journal of Dynamic Systems, Measurement, and
Control. 138.
Leite Gustavo de Novaes Pires, Araújo Alex Maurício & Rosas Pedro André Carvalho. 2018.
Prognostic techniques applied to maintenance of wind turbines: a concise and specific re-
view. Renewable and Sustainable Energy Reviews, Vol. 81, Part 2, pp. 1917–1925.
Leturiondo Urko. 2016. Hybrid Modelling in Condition Monitoring. Doctoral thesis. Luleå
University of Technology. Operation and Maintenance. ISBN 978-91-7583-722-2.
MacGregor John & Cinar Ali. 2012. Monitoring, fault diagnosis, fault-tolerant control and
optimization: Data driven methods. Computers and Chemical Engineering. Vol. 47, pp. 111–
120.
Manning-Ohren David. 2015. Top ten rules for power plant condition monitoring. Power
Engineering International [Online]. [Accessed on 28.9.2017]. Available: http://www.pow-
erengineeringint.com/articles/print/volume-23/issue-5/features/top-ten-rules-for-condition-
monitoring.html
Mansouri Majdi, Mohammed Z.S., Baklouti Raoudha, Nounou Mohamed, Nounou Hazem,
Hamida Ahmed Ben & Karim Nazmul. 2016. Statistical fault detection of chemical process
– comparative studies. Journal of Chemical Engineering & Process Technology, Vol. 7, Iss.
1.
MANTIS. 2015. About MANTIS. [Online]. [Accessed on 26.1.2018]. Available:
http://www.mantis-project.eu/about-mantis/.
Matignon Randall. 2005. Neural Network Modeling Using Sas Enterprise Miner. Autho-
rHouse. ISBN 1-4184-6675-1.
Marlow D.O., Sanchez S.M. & Sanchez P.J. 2017. Assessing the impact of the bathtub curve
failure rate on fleet performance using designed simulation experiments. 22nd International
Congress on Modelling and Simulation, Hobart, Tasmania, Australia, 3 to 8 December 2017.
Martha de Souza Gilberto Francisco. 2012. Thermal Power Plant Performance Analysis.
Dordrecht: Springer. ISBN 978-1-4471-2309-5.
Marwala Tshilidzi. 2012. Condition Monitoring Using Computational Intelligence Methods.
London: Springer. ISBN 978-1-4471-2379-8. J. Chem. Eng. Process Technol. Vol. 7, no 1,
pp. 1–10.
May Robert, Dandy Graeme, Maier Holger & Suzuki Kenji (Ed.). 2011. Review of Input
Variable Selection Methods for Artificial Neural Networks – Methodological Advances and
Biomedical Applications. InTech. ISBN: 978-953-307-243-2.
Mendel Jerry M. 1995. Fuzzy Logic Systems for Engineering: A Tutorial. Proceedings of
the IEEE. Vol. 83(3), pp. 345–377.
Mobley Keith R. 2002. An Introduction to Predictive Maintenance. 2nd Edition. Elsevier
Science. ISBN 0-7506-7531-4.
Moubray John. 1997. Reliability-centered Maintenance. 2nd Edition. New York: Industrial
Press Inc. ISBN 0-8311-3078-4.
Ndjenja L. & Visser J.K. 2015. Development of a Maintenance Strategy for Power Genera-
tion Plants. South African Institute of Electrical Engineers. Vol. 106, pp. 132–140.
Niu Gang. 2017. Data-Driven Technology for Engineering Systems Health Management.
Beijing: Springer. ISBN 978-981-10-2031-5.
Niu Gang, Yang Bo-Suk & Pecht Michael. 2010. Development of an optimized condition-
based maintenance system by data fusion and reliability-centered maintenance. Reliability
Engineering and System Safety. Vol 95, pp. 786–796.
Opetushallitus. Värähtelymittaukset [Online]. [Accessed on 20.9.2017]. Available:
http://www03.edu.fi/oppimateriaalit/kunnossapito/mekaniikka_k2_varahtelymit-
taukset.html
Patan Krzysztof. 2008. Artificial Neural Networks for the Modelling and Fault Diagnosis of
Technical Processes. Verlag: Springer. ISBN 978-3-540-79871-2.
Pecht Michael & Gu Jie. 2009. Physics-of-failure-based prognostics for electronic products.
Transactions of the Institute of Measurement and Control, Vol. 31, Iss. 3–4, pp. 309–322.
Pecht Michael & Jaai Rubyca. 2010. A prognostics and health management roadmap for
information and electronics-rich systems. Microelectronics Reliability 50, pp. 317–323.
Pusey Henry C. 2007. Turbomachinery Condition Monitoring and Failure Prognosis. Sound
And Vibration March 2007, pp. 8–12.
Qin Joe S. 2012. Survey on data-driven industrial process monitoring and diagnosis. Annual
Reviews in Control. Vol. 36, pp. 220–234.
Qiu Tian & Bai Xiaojing. 2012. Comparison of Different Fault Detection Statistics Detect-
ability in PCA. The 2nd International Conference on Computer Application and System Mod-
eling.
Radhakrishnan Pranoy. 2017. Introduction to Recurrent Neural Network. Towards Data Sci-
ence. [Online]. [Accessed on 14.3.2018]. Available: https://towardsdatascience.com/intro-
duction-to-recurrent-neural-network-27202c3945f3
Rasila Antti. 2007. Sovellettu todennäköisyyslaskenta B. Aalto University. [Online]. [Ac-
cessed on 18.3.2018]. Available: http://www.mv.helsinki.fi/home/mmattila/kvanti/lu-
ento8.pdf
Reisinger Heribert. 1997. The impact of research designs on R2 in linear regression models:
an exploratory meta-analysis. Journal of Empirical Generalisations in Marketing Science,
Vol. 2.
Roulias Dimitrios, Louats Theodoros & Kostopoulos Vassilios. 2013. A Statistical Feature
utilising Wavelet Denoising and Neighblock Method for Improved Condition Monitoring of
Rolling Bearings. Chemical Engineering Transactions. Vol. 3, pp. 1045–1050.
Severson Kristen, Chaiwaranodom Paphonwit & Braatz Richard D. 2016. Perspectives on
process monitoring of industrial systems. Annual Reviews in Control. Vol. 42, pp. 190–200.
Shanmuganathan Subana. 2016. Artificial Neural Network Modelling: An Introduction.
Book Section in: Shanmuganathan Subana (edit) & Sandhya Samarasinghe (edit). Artificial
Neural Network Modelling, pp 1–14. Switzerland: Springer. ISBN 978-3-319-28493-4.
Sheriff Mohammed, Botre Chiranjivi, Mansouri Majdi, Nounou Hazem, Nounou Mohamed
& Karim Mohammad. 2017. Process Monitoring Using Data-Based Fault Detection Tech-
niques: Comparative Studies. Book Section in: Demetgul Mustafa (edit.) & Ünal Muhammet
(edit.). Fault Diagnosis and Detection, pp. 238-261. InTech. ISBN 978-953-51-3204-2.
Shin Jong-Ho & Jun Hong-Bae. 2015. On condition based maintenance policy. Journal of
Computational Design and Engineering, Vol. 2, pp. 119–127.
Simani S. & Fantucci C. 2000. Fault diagnosis in power plant using neural networks. Infor-
mation Sciences. Vol. 127, Iss. 3–4, pp. 125–136.
Singh Amarjit & Adachi Stacy. 2013. Bathtub curves and pipe prioritization based on failure
rate. Built Environment Project and Asset Management, Vol. 3 Iss. 1, pp.105-122.
Sobhani-Tehrani Ehsan & Khorasani Khashayar. 2009. Fault Diagnosis of Nonlinear Sys-
tems Using a Hybrid Approach. Springer: Dordrecht. ISBN 978-0-387-92906-4.
Song Yong-Hua & Johns Allan T. 1997. Applications of fuzzy logic in power systems. I.
General introduction to fuzzy logic. Power Engineering Journal. Vol. 11, Iss. 5, pp. 219–
222.
Stanley Greg. 2010. Model Based Reasoning for Fault Detection and Fault Diagnosis.
[Online]. [Accessed on 30.10.2017], Available: https://gregstanleyandassociates.com/white-
papers/FaultDiagnosis/Model-Based-Reasoning/model-based-reasoning.htm
Stefaniak Pawel, Zimrow Radoslaw, Sliwinski Pawel, Andrzejewski Marek & Wylomanska
Agnieszka. 2014. Multidimensional Signal Analysis for Technical Condition, Operation and
Performance Understanding of Heavy Duty Mining Machines. In: Chaari Fakher, Zimroz
Radoslaw, Bartelmus Walter & Haddar Mohammed (editors). Advances in Condition Mon-
itoring Machinery in Non-Stationary Operations. Cham: Springer. pp. 197–210. ISBN 978-
3-319-20462-8.
Stringer David, Sheth Pradip & Allaire Paul. 2012. Physics-based modeling strategies for
diagnostic and prognostic application in aerospace systems. Journal of Intelligent Manufac-
turing, Vol. 23, Issue 2, pp. 155–162.
Suganthi L., Inyan S. & Samuel Anand A. 2015. Applications of fuzzy logic in renewable
energy systems – A review. Renewable and Sustainable Energy Reviews. Vol. 46, pp. 585–
607.
Sullivan G.P., Pugh R., Melendez A.P. & Hunt W.D. 2010. Operations and Maintenance
Best Practices Guide, Release 3.0. U.S. Department of Energy. Federal Energy Management
Program.
Sutharssan Thamo, Stoyanov Stoyan, Bailey Chirs & Yin Chunyan. 2015. Prognostic and
health management for engineering systems: a review of the data-driven approach and algo-
rithms. The Journal of Engineering. July 2015.
Tahan Mohammadreza, Tsoutsanis Elias, Muhammad Masdi & Karim Z.A. Abdul. 2017.
Performance-based health monitoring, diagnostics and prognostics for condition-based
maintenance of gas turbines: A review. Applied Energy, Vol. 198, pp. 122–144.
Talebi Heidar A., Abdollahi Farzaneh, Patel Rajni V. & Khorasani Khashayar. 2010. Neural
Network-Based State Estimation of Nonlinear Systems – Application to Fault Detection and
Isolation. New York: Springer. ISBN 978-1-4419-1437-8.
Tautz-Weinert J. & Watson S.J. 2017. Combining Model-based Monitoring and a Physics
of Failure Approach for Wind Turbine Failure Detection. Presented at the 30th Conference
on Condition Monitoring and Diagnostic Engineering Management (COMADEM 2017),
University of Central Lancashire, UK, 10-13th July.
Temsamani A.B., Kauffmann S., Descas Y., Vandevelde B., Zanon F. & Willems G. 2017.
Improved and accurate physics-of-failure (PoF) methodology for qualification and lifetime
assessment of electronic systems. Microelectronics Reliability. Vol. 76–77, pp. 42–46.
Tetko Igor V., Livingstone David J. & Luik Alexander I. Neural Network Studies. 1. Com-
parison of Overfitting and Overtraining. J: Chem. Inf. Comput. Sci., Vol. 35, No 5, pp. 826–
833.
The World Energy Council, 2016. World Energy Recourses 2016. ISBN 978 0 946121 58
8.
Thurston Michael & Lebold M. 2001. Open Standards for Condition-Based Maintenance
and Prognostic Systems. 5th Annual Maintenance and Reliability Conference (MARCON
2001), Galtinburg, TN, May 25–27.
Tidriri Khaoula, Chatti Nizar, Verron Sylvain & Tiplica Teodor. 2016. Bridging data-driven
and model-based approaches for process fault diagnosis and health monitoring: A review of
researches and future challenges. Annual Reviews in Control. Vol. 42, pp. 63–81.
Twomey Janet M. & Smith Alice E. Validation and Verification. Chapter in: Kartam Nabil,
Flood Ian & Garrer James H. 1997. Arftificial Neural Networks for Civil Engineers: Funda-
mentals and Applications. New York: ASCE Press. ISBN 0784402256.
Tzafestas S.G., Dalianis P.J. % Anthopoulos G. 1996. On the overtraining phenomenon of
backpropagation neural networks. Mathematics and Computers in Simulation. Vol 40, pp.
507–521.
Urata Mikiyasu, Aoki Kiyotaka, Urata Naoya, Tagami Koichi, Kondo Manabu & Matsu-
moto Shintaro. 2015. Non Destructive Inspection Technology for Thermal Power Plants.
Mitsubishi Heavy Industries Technical Review. Vol. 52 No. 4, pp. 37–42.
Van Tung Tran & Yang Bo-Suk. 2009. Machine Fault Diagnosis and Prognosis: The State
of The Art. International Journal of Fluid Machinery and Systems. Vol. 2, pp. 61–71.
Venkatasubramanian Venkat, Rengaswamy Raghunathah, Yin Kevin & Kavuri Surya.
2003a. A review of process fault detection and diagnosis: Part I: Quantitative model-based
methods. Computers and Chemical Engineering. Vol 27, Iss. 3, pp. 293–311.
Venkatasubramanian Venkat, Rengaswamy Raghunathah, Yin Kevin & Kavuri Surya.
2003b. A review of process fault detection and diagnosis: Part III: Process history based
methods. Computers and Chemical Engineering. Vol 27, Iss. 3, pp. 327–346.
Villegas Thamara, Fuente Maria Jesus & Rodriguez Miguel. Principal Component Analysis
for Fault Detection and Diagnosis. Experience with a pilot plant. International Confer-
ence on Computational Intelligence, Man-Machine Systems and Cybernetics - Proceedings
2010. pp. 147–152.
Westerhuis Johan, Gurden Stephen & Smilde Age. 2000. Generalized contribution plots in
multivariate statistical process monitoring. Chemometrics and Intelligent Laboratory Sys-
tems. Vol. 51, Iss. 1, pp. 95–114.
Wireman Terry. 2004. Total Productive Maintenance. 2nd Edition. New York: Industrial
Press Inc. ISBN 0-8311-3172-1.
Yan Jihong. 2014. Machinery Prognostics and Prognosis Oriented Maintenance Manage-
ment. Singapore: John Wiley & Sons. ISBN 111863876X.
Yang Qingsong. 2004. Model-Based and Data Driven Fault Diagnosis Methods with Appli-
cations to Process Monitoring. Doctoral thesis. Case Western Reserve University. Electrical
Engineering and Computer Sciences.
Zhang Q.J. & Gupta K.C. 2000. Neural Networks for RF and Microwave Design. Artech
House, Inc. Norwood, MA, USA. ISBN 1580531008.
Zhang Wan, Jina Min-Ping, Zhu Lin & Yan Xiao-An. 2017. Comprehensive Overview on
Computational Intelligence Techniques for Machinery Condition Monitoring and Fault Di-
agnosis. Chinese Journal of Mechanical Engineering. Vol. 30 (4), pp. 782–795.
Zhang Zhenyou, Wang Yi & Wang Kesheng. 2013. Intelligent fault diagnosis and prognosis
approach for rotating machinery integrating wavelet transform, principal component analy-
sis, and artificial neural networks. The International Journal of Advanced Manufacturing
Technology. Vol. 68, Iss. 1–4, pp. 762–773.
Zhou Shaoyuan, Zhang Jianming & Wang Shuquing. 2004. Fault Diagnosis in Industrial
Processes Using Principal Component Analysis and Hidden Markov Model. Proceedings of
the 2004 American Control Conference. Vol. 6.
Zhu Shun-Peng, Huang Hong-Zhong, Peng Weiwen, Wang Hai-Kun & Mahadevan Sanka-
ran. 2016. Probabilistic Physics of Failure-based framework for fatigue life prediction of
aircraft gas turbine discs under uncertainty. Reliability Engineering & System Safety. Vol.
146, pp. 1–12.
